Edge Deep Learning in Computer Vision and Medical Diagnostics: A Comprehensive Survey

• Edge computing processes data near its source, complementing cloud computing while reducing latency and supporting privacy, scalability, distributed processing, and near-real-time applications.Source: §1 · p. 2
• Modern deep-learning models can exceed edge-device resources, motivating lightweight designs and model compression for efficient deployment.Source: §1 · p. 2
• Edge DL combines local computation with AI across many domains, and in medical diagnostics can provide immediate image-analysis and monitoring feedback to improve diagnostic efficiency and patient safety.Source: §1 · pp. 2–3
• The survey covers edge concepts and hardware, model compression and lightweight design, computer-vision and medical applications, and research challenges and future opportunities.It presents a comprehensive reference focused on computer vision generally and medical diagnostics particularly.Source: §1 · p. 3

2.1 Edge Computing Related Paradigms and Terminologies

• Cloudlets provide nearby mini-cloud services, fog computing distributes processing across network-edge nodes, and MEC brings computation close to mobile devices for lower latency and bandwidth use.Cloudlets bridge edge devices and large cloud data centres; fog computing complements them with distributed processing, storage, and analysis.Source: §2.1 · p. 5
• Edge deep learning and edge analytics process or train models near data sources, while cloud-edge collaboration combines resources; EdgeMove, federated learning, and distributed training support edge model training (Wang et al. 2020; Yao et al. 2022; Dong et al. 2023; Brecko et al. 2022; Abreha et al. 2022; Hu et al. 2020).Federated learning trains on local data while preserving privacy, but requires computational and communication resources.Source: §2.1 · p. 6

2.2 Advantages of Edge Computing

• Local processing supports rapid ambulance, ICU, and MRI/CT diagnosis, potentially accelerating treatment planning in time-sensitive care (Sait et al. 2019; Liu et al. 2023; Chen et al. 2019; Lingappa and Parvathy 2022).The survey also identifies reduced transmission and storage costs and lower cloud-computing burdens as benefits of processing data locally.Source: §2 · p. 4, §2.2 · p. 7
• Local processing reduces exposure of sensitive health data during transmission and can avoid transferring large medical images, conserving bandwidth (Cao et al. 2023; Singh and Chatterjee 2021; Alwakeel 2021; Dong et al. 2020; Dave et al. 2021; Zheng et al. 2021).The survey highlights local handling of patient records and diagnostic images as particularly relevant to healthcare privacy and medical imaging.Source: §2.2 · pp. 7–8
• Edge computing supports scalable remote care and continuity despite unreliable connectivity, while local processing and storage, cloud integration, and cheaper hardware can improve healthcare cost efficiency.These capabilities support continuous monitoring and care in remote or rural settings and can reduce cloud storage and data-transmission costs.Source: §2.2 · pp. 8–9

2.3 Disadvantages of Edge Computing

• Medical edge deployment is constrained by limited compute for large scans, heterogeneous hardware and software, and inconsistent results without standards, complicating accuracy, compatibility, compliance, and maintenance.Compression or lightweight models may be needed, but can compromise diagnostic accuracy; standardised protocols help ensure consistency and regulatory compliance.Source: §2.3 · pp. 9–10

3.1 Edge Computing Device Categories

• Low-end, medium-range, and high-performance devices span basic preprocessing and inference through complex model training, enabling selection by computational needs and application requirements.Low-end and medium-range devices are constrained in training but can provide near-real-time inference; local processing also reduces data transmission and its associated privacy and security risks.Source: §3 · p. 10, §3.1 · pp. 10, 12, 13
• Low-end devices favor energy efficiency and affordability for basic sensing and inference, medium-range devices support more advanced inference, and high-performance devices enable training and complex medical-image analysis.The categories range from constrained memory and processors to powerful CPUs, GPUs, and substantial memory; high-performance devices can locally train models using test-time domain adaptation and continual learning.Source: §3.1 · pp. 10–12
• Portable low-end and medium-range devices suit space-constrained or mobile deployments, while local near-real-time inference can reduce data transmission and risks of leakage or tampering.The section notes these devices’ limited training capacity alongside their portability and local processing advantages.Source: §3.1 · pp. 12–13

3.2 Detailed Overview of Edge Computing Devices

• The overview spans microcontrollers, accelerator boards, smartphones, and workstations, underscoring that device choice should match computational and application requirements.Examples include low-power embedded platforms, mid-range devices for real-time inference, and configurable workstations for resource-intensive training.Source: §3.2 · pp. 13, 17
• The YOLOv3 comparison reports Jetson Xavier NX advantages in FPS and inference time, while Jetson Nano balances performance and energy efficiency (Feng et al. 2022).The comparison used 1,596 video frames at 768×436 resolution; the passage supplies no specific metric values.Source: §3.2 · p. 17, Table 3 · p. 17

4 Transferring Large DNNs to Lightweight DNNs

• Lightweight neural networks suit resource-limited edge devices, offering smaller memory footprints, faster inference, and lower energy consumption than larger counterparts (Khan et al. 2022; Javed et al. 2024; Naqvi et al. 2023).The passage also cites Khan et al. 2022, 2024 and Farooq et al. 2024 on the development of lightweight architectures.Source: §4 · p. 18
• The section examines network pruning, sparse representation, quantised neural networks, efficient lightweight neural networks, and neural architecture search as routes to lightweight models for edge devices.Figure 4 presents the methods; Table 4 compares their advantages, disadvantages, and potential applications.Source: §4 · p. 18
• Lightweight architectures aim to balance computational efficiency with model complexity for implementation on edge devices.Source: §4 · p. 18

4.1.1 Weight Pruning

• Unstructured weight pruning reduces complexity by 9× in AlexNet and 13× in VGG16 through threshold-based weight removal, retraining, and sparse-matrix storage.Pruning sets selected weights to zero; the sparse matrix is stored using compressed sparse row or column structure.Source: §4.1.1 · p. 19

4.1.2 Neuron Pruning

• Neuron pruning removes entire neurons with their associated weights, and structured pruning can target neurons with frequent zero activations after ReLU, measured using APoZ.The model is typically pruned iteratively and fine-tuned to recover performance; the passage also notes that redundancy can worsen overfitting.Source: §4.1.2 · p. 20

4.1.3 Channel Pruning

• Channel pruning creates structured sparsity by reducing feature-map channels, enabling CPU and GPU implementation; inference-based pruning selects representative channels and rebuilds feature maps using linear least squares.This approach addresses the difficulty of implementing sparse weight-pruned models on hardware without specialized inference support.Source: §4.1.3 · p. 20
• Channel pruning of MobileNetV1 achieved a 2× reduction in FLOPs while maintaining 70.9% top-1 accuracy on ImageNet, and its structured sparsity supports implementation across different hardware without specialized inference engines.Source: §4.1.3 · p. 21
• Learning-based channel pruning reduces ResNet-50 parameters by 60.8% while maintaining 98% of original ImageNet accuracy; channel pruning with quantisation yields 69.86% top-1 accuracy and 5× smaller ResNet-34.The combined channel-pruning and quantisation result is reported for ImageNet.Source: §4.1.3 · p. 21

4.1.4 Filter Pruning

• Filter pruning removes less impactful filters, whose importance can be assessed using weight magnitudes, activation patterns, or gradient information.The method is presented as a way to streamline models for resource-constrained settings such as edge devices and mobile applications.Source: §4.1.4 · p. 21
• Filter pruning accelerates ResNet-50 by 2× while maintaining 98.6% of original ImageNet accuracy; MobileNetV2 pruning reduces FLOPs by 2× while maintaining 70.6% top-1 accuracy.The MobileNetV2 result uses ImageNet.Source: §4.1.4 · p. 22
• Learning-based filter pruning reduces ResNet-56 FLOPs by 60.3% while maintaining 99% of original CIFAR-10 accuracy; filter pruning with quantisation yields 74.69% top-1 accuracy and a 5.3× smaller ResNet-101.The combined filter-pruning and quantisation result is reported for ImageNet.Source: §4.1.4 · p. 23

4.2 Sparse Representation

• L1 regularisation encourages parameters to become zero during training, while ReLU can naturally produce sparse activations.Both are described as techniques for inducing sparsity in neural networks.Source: §4.2 · p. 23
• Sparse networks reduce computational and memory requirements, can improve interpretability, and act as regularisation to prevent overfitting and improve robustness.Nonzero weights and activations can be associated with influential features and connections.Source: §4.2 · p. 23
• Selecting an optimal sparsity level and managing performance trade-offs remain challenges, prompting work on pruning, adaptive sparsity, and mitigation strategies.Sanh et al. (2020) used movement pruning to dynamically modify BERT sparsity during training, reducing model size while maintaining much of its performance on SQuAD, MNLI, and QQP.Source: §4.2 · pp. 23–24
• Qi et al. (2021) used structured pruning on ResNet-50 to reduce parameters 3× and FLOPs 2× while retaining 97% of original ImageNet accuracy.The approach prunes entire filters or layers and is presented as suitable for real-time image classification on edge devices.Source: §4.2 · p. 24
• Hu et al. (2023) combined quantisation and sparsity on MobileNetV2, achieving 73.1% top-1 accuracy on ImageNet with reduced model size and inference latency.The passage identifies mobile-device image recognition and augmented reality as potential applications.Source: §4.2 · p. 24

4.3 Quantised Neural Network

• Quantisation maps high-precision parameters to lower-bit values, including 8-bit weights and activations or 1-bit binary representations, while mixed precision assigns bit-widths by layer sensitivity.Lower bit-widths can reduce memory requirements, inference time, computation, and energy use; Fig. 8 illustrates mapping floating-point parameters to integers.Source: §4.3 · pp. 24–25, Figure 8 · p. 26
• Reduced precision can lower accuracy, especially on fine-grained tasks, so quantisation-aware training and task- and platform-appropriate bit-width selection are important.The passage cites Chung et al. (2020) and Shen et al. (2021) for quantisation-aware training.Source: §4.3 · p. 25
• Adaptive quantisation can change bit-widths during training based on weight and activation distributions, while learned intervals adjust to data characteristics rather than remaining fixed.The passage attributes adaptive quantisation to Dong et al. (2019) and learned intervals to Lee and Kim (2018), describing potential improvements for intricate tasks and model reliability and precision.Source: §4.3 · p. 25
• ResNet-50 reached 75. 3% top–1 accuracy on ImageNet with mixed-precision quantisation, reducing model size by 10× and inference latency versus full precision; quantised BERT cut memory 4× while retaining 98% of its original GLUE accuracy.Wang et al. (2019) report the ResNet-50 result; Zafrir et al. (2019) report the BERT result and describe it as enabling real-time mobile text processing.Source: §4.3 · p. 25
• MobileNetV2 with hardware-aware quantisation achieved 72.2% top-1 ImageNet accuracy at 220 million FLOPs, with significant energy savings for power-limited edge devices.The result is attributed to Huang et al. (2024).Source: §4.3 · p. 26

4.4.1 MobileNet

• MobileNetV3-Large achieved 75.2% top-1 accuracy on ImageNet with 219 million MAdds, while MobileNetV2 pneumonia detection reached 96.08% precision and 96.71% accuracy.MobileNetV1 introduced depthwise separable convolutions, V2 added inverted residual blocks and linear bottlenecks, and V3 added h-swish; variants also support smartphone-based plant disease detection, wearable ECG monitoring, and mobile robotics. MobileNetV1 recorded 70.6% top-1 accuracy with 569 million FLOPs, and MobileNetV2 recorded 72.0% with 300 million MAdds.Source: §4.4.1 · pp. 26–27

4.4.2 EfficientNet

• EfficientNet-B0 achieved 77.1% top-1 accuracy on ImageNet with 5.3 million parameters and 390 million FLOPs, while EfficientNetV2-S achieved 83.9% top-1 accuracy.EfficientNet uses compound scaling across width, depth, and image resolution; larger variants trade increased capacity and computation for performance. EfficientNet in diabetic-retinopathy detection attained 98.1% sensitivity and 98.3% specificity while reducing inference time.Source: §4.4.2 · pp. 27–28

4.4.3 Group Convolutional Networks

• ShuffleNetV2 achieved 69.4% top-1 accuracy on ImageNet with 146 million FLOPs, while CondenseNet achieved 73.8% with 4.8 million parameters.ShuffleNet uses group convolutions and channel shuffling, whereas CondenseNet prunes and condenses less important connections; both are presented as efficient options for resource-constrained devices.Source: §4.4.3 · p. 29
• MixNet-M achieved 77.0% top-1 accuracy on ImageNet with 360 million FLOPs, GhostNet achieved 73.9% with 141 million FLOPs, and DiCENet achieved 75.7% top-1 accuracy.MixNet uses mixed depthwise convolutions, GhostNet generates feature maps with ghost modules, and DiCENet dynamically adjusts active inference channels; MicroNet targets ultra-low-resource edge and IoT use and can match or surpass MCUNet and EfficientNet-Lite.Source: §4.4.3 · pp. 29–30

4.4.4 Squeeze & Excitation

• SqueezeNet reduced parameters 50× and inference time 2× versus AlexNet, SENet improved MobileNetV2 top-1 accuracy from 70.6% to 74.7%, and SqueezeNeXt-23 reached 67.18%.SqueezeNet uses 1×1 convolutions in fire modules; SENet recalibrates channels through squeeze-and-excitation attention; SqueezeNeXt combines that attention with parallel paths of varying channel sizes and recorded 2.4 million parameters and 1.5 billion FLOPs.Source: §4.4.4 · pp. 31–32

4.4.5 Mobile Transformers

• MobileViT reached 78.4% top-1 accuracy on ImageNet-1k, while EdgeViT-S reached 81.0% top-1 accuracy at 85.3±3.9 ms on Samsung Galaxy S21.MobileViT surpassed MobileNetV3 by 3.2% and DeiT by 6.2% on ImageNet-1k, and was 5.7% more accurate than MobileNetV3 on MS-COCO; EdgeViT-S uses a local-global-local bottleneck, while mobile transformers are also applied to facial recognition, medical imaging, retail, and AR. MobileViT combines lightweight CNNs with self-attention; EdgeViT-S targets low latency and energy use.Source: §4.4.5 · pp. 32–33

4.5 Neural Architecture Search

• NAS seeks task-specific architectures through automated search, reducing reliance on human expertise and manual trial and error.The section briefly reviews notable NAS approaches and architectures, while directing readers to Elsken et al. (2019) for a comprehensive survey.Source: §4.5 · p. 33
• PNASNet’s progressive search achieved a 3.41% top-1 error rate on CIFAR-10 using 400 GPU hours, versus NASNet’s 2,000 GPU hours.Its progressive strategy grows network complexity incrementally to keep search computationally feasible; the passage also reports 82.9% top-1 accuracy on CIFAR-10.Source: §4.5 · pp. 33–34
• MNASNet’s latency-aware search achieved 75.2% top-1 accuracy on ImageNet with 312 billion MAdds, while FBNet achieved 74.9% with 375 million FLOPs (Tan et al. 2019; Wu et al. 2019).MNASNet targets mobile and edge devices through multi-objective optimization incorporating latency; FBNet factorizes operation and connectivity selection and is reported as 1.5× more efficient than MobileNetV2.Source: §4.5 · p. 34
• AmoebaNet used evolutionary search and achieved 96.7% top-1 accuracy on CIFAR-10 and 83.9% on ImageNet (Real et al. 2019a).The passage also reports that its 83.9% ImageNet result was state of the art at the time.Source: §4.5 · p. 34
• NAS supported edge applications including real-time drone navigation, wearable physiological monitoring, and automotive ADAS and autonomous-driving models (Cereda et al. 2023).The examples address cloud-independent real-time operation, low-power monitoring through quantization and pruning, and vehicle latency and energy-efficiency requirements.Source: §4.5 · pp. 34–35

5.1 Autonomous Driving and Intelligent Transportation Systems

• A YOLOv4 method modified in its backbone, neck, and predictor head improved detection accuracy and inference speed, with higher average accuracy on KITTI and BDD for real-time autonomous-driving applications.Liu et al. (2023) used RGB-camera–LiDAR fusion to improve detection performance, particularly in complex and dynamic driving environments; Wan et al. (2022), Fernández et al. (2021), and Zou et al. (2022) describe edge-based traffic monitoring or surveillance approaches.Source: §5.1 · pp. 36–37

5.2 Smart Manufacturing and Industrial Automation

• At least 22.9% faster detection, 35.0% higher average precision, and 8.4% lower memory consumption were reported for lightweight ceramic-defect detection versus classical methods.Chen et al. (2022) report 121 frames/second and an mAP of 89.8%, with a 44% speed increase over standard YOLOv3; Bonam et al. (2023) report test accuracy of 96.87% on fabric, 99% on surface, and 98.08% on casting defects using lightweight models deployed on Jetson Nano. Methods also include lightweight semantic segmentation and YOLOv5-based defect detection; legacy-system compatibility and initial deployment costs remain challenges.Source: §5.2 · pp. 37–38

5.3 Agriculture and Crop Monitoring

• UAV-based edge vision supports pest detection, irrigation assessment, and crop phenotyping, with performance affected by crop type and environmental conditions.In a Cabernet Sauvignon orchard, UAV imagery found significant yield variation from deficient irrigation and correlations between NDVI, GNDVI, canopy temperature, fruit yield, and stomatal conductance; citrus-tree detection and counting achieved 99.9% precision and 99.7% recall across 4,931 trees.Source: §5.3 · pp. 39–40

5.4 Retail and Shelf Monitoring

• Retail edge vision automates shelf and inventory monitoring, from low-power FPGA shelf-status cameras to mobile robots that detect out-of-stock products.CMSS consumes below 6 mW and has a maximum lateral recognition distance of 40 centimeters; ROSCH, using a Pepper robot and YOLOv6, detected out-of-stock items eight times faster than manual detection. Retail systems may require substantial initial infrastructure investment.Source: §5.4 · pp. 41–42

5.5 Environmental Monitoring and Wildlife Conservation

• DeepWILD detected, classified, and counted 13 Parc National du Mercantour species under various environmental conditions, achieving 96.88% mAP at IoU 0.5; another edge approach used wireless updates and an active-learning pipeline.Gupta et al. (2022) used self-supervised CNNs to address cluttered camera-trap images associated with low detection and high false-discovery rates.Source: §5.5 · pp. 43–44
• Verma and Gupta (2018) used transfer-learning models for elephant detection, while wildlife monitoring also raises concerns about disturbance and animal behaviour.Their visual-cue early-warning system aimed to prevent human-elephant collisions; noninvasive camera traps can minimise human disturbance and potential stress or behaviour changes.Source: §5.5 · p. 44

5.6 Smart Cities and Public Safety

• Smart-city edge vision supports real-time crowd and safety monitoring, including people counting, PPE checks, occluded-face recognition, and fire detection.Avvenuti et al. (2022) used temporal attention to reduce FDST counting errors; Di Benedetto et al. (2022) evaluated an embedded safety system on two datasets. Deng et al. (2023) addressed occluded faces with grayscale inputs, while Muhammad et al. (2018) optimised fire detection with MobileNet for faster, more efficient inference.Source: §5.6 · pp. 45–46
• Smart-city vision applications also include drone-based assistance for visually impaired people and lightweight small-object detection for urban disaster response.Raj et al. (2023) proposed drones for navigation, hazard detection, and collision avoidance; Cao et al. (2023) combined MobileNetV3, ECA attention, and extra prediction heads for drone imagery.Source: §5.6 · pp. 45–46
• Chen et al. (2006) used satellite imagery and the Normalised Difference Bareness Index to examine Shenzhen’s urban heat island and inform urban planning.Their analysis related land use/cover and brightness temperature to urban expansion and regional climate, with potential to support mitigation of urban heat effects.Source: §5.6 · pp. 45–46

5.7 Sports Analytics

• Sports studies apply edge or distributed vision to video processing, action localisation, and safer Taekwondo training.Cui and Hu (2022) proposed federated learning and a lightweight distributed model to reduce video transmission load and latency, and explored sensor-assisted action recognition. Cioppa et al. (2020) introduced a temporal-context loss that improved action discovery over a baseline on SoccerNet and was applied to ActivityNet.Source: §5.7 · pp. 46–47
• Sports vision research also advances motion analysis and camera calibration using 2D video and learned segmentation.Baumgartner and Klatt (2023) used partial motion field registration and a synthetic Unreal Engine 5 dataset for 3D pose estimation; Van Hooren et al. (2023) evaluated DeepLabCut and OpenPose against marker-based running kinematics. Zhang and Izquierdo (2023) used cGAN-generated segmentation and regression to estimate four camera-calibration points for real-time use.Source: §5.7 · p. 47

6 Medical Edge Applications

• Edge hardware such as NVIDIA Jetson devices and Raspberry Pi, alongside smartphones and lightweight networks, supports low-power, real-time medical diagnosis.Examples include Jetson-based skin and cervical cancer detection and Raspberry Pi or smartphone approaches for chest CT and dermatology. The survey also covers gastrointestinal, pulmonary and thoracic, dermatological, pathological-image, and telemedicine applications.Source: §6 · pp. 47–48

6.2 Pulmonary and Thoracic Diagnosis

• Edge pulmonary and thoracic systems support CXR and CT analysis through embedded, mobile, and clinical devices, including lightweight models and secure browser-based imaging.Masud et al. (2020) report 97.9% accuracy for a lightweight mobile CNN, while Raghavan et al. (2020) report lung cancer detection rates comparable to the National Lung Screening Trial, including for patients not eligible for medical insurance.Source: §6.2 · pp. 51–52

6.3 Dermatological Diagnosis

• Dermatological edge applications use mobile and compact devices for timely skin-disease assessment, including local processing, rural prescreening, and lesion classification.Shrivastava et al. (2023) report up to 93.3% classification accuracy on the PH2 skin lesion dataset using a Jetson Nano-targeted model. Dai et al. (2019) perform computations locally to reduce latency and bandwidth use while enhancing privacy; Ngeh et al. (2020) describe internet-independent prescreening with a Raspberry Pi.Source: §6.3 · pp. 53–54

6.4 Pathological Images Analysis

• Pathology applications bring digital tissue-slide analysis to edge settings, addressing complex, time-consuming interpretation with lightweight models and portable viewing or microscopy.Ramey et al. (2011) found 89% concordance between initial frozen-section diagnoses and iPad diagnoses, with an average of less than 2 minutes per slide. Auguste and Palsana (2015) describe a portable, economical whole-slide imaging method using a microscope and iPhone.Source: §6.4 · pp. 54–55
• Histopathology edge systems extend beyond tissue-slide classification to embedded cervical-cancer detection and real-time cell imaging and monitoring.The survey characterizes histopathology as the gold standard for cancer diagnosis and prognosis, while noting that slide interpretation is time-consuming and subject to inter- and intra-observer variability.Source: §6.4 · pp. 54–55
• ReducedFireNet achieved 96.88% average accuracy and an F1 score of 0.968 on Malignant Lymphoma datasets; Wang et al. (2019) reported 94.1% accuracy on Jetson TK1.ReducedFireNet measures 0.391 MB and 0.201 GFLOPS. Wang et al.'s network uses optimized convolutions, parameter compression, and enhanced feature expression depth.Source: §6.4 · p. 55
• CytoPAN,58, a portable image cytometer, is reported to show exceptional diagnostic precision for breast cancer subtypes, especially ER/PR and HER2.The passage also describes cell analysis as accurate and reliable.Source: §6.4 · p. 56

6.5 Telemedicine

• FootSnap showed high intra- and inter-operator reliability for longitudinal DFU monitoring, with Jaccard similarity index values of 0.89 to 0.94.Earlier iPhone-based assessments compared image diagnoses with on-site podiatrist assessments; UTrack adds mobile wound capture, segmentation, measurement, monitoring, and local data storage for privacy.Source: §6.5 · p. 56
• 91.8% average precision and 48 ms per image were achieved by a transfer-learned Fast R-CNN method for diabetic-foot-ulcer detection on an NVIDIA Jetson TX2.The method used 1,775 expert-annotated DFU images.Source: §6.5 · p. 56
• Edge-based fall monitoring uses mobile IoT devices to reduce data transmission and storage needs, while supporting timely assistance and elderly self-sufficiency.A separate 3D human-pose-estimation app on the NVIDIA Jetson Xavier enables markerless, real-time remote gait analysis.Source: §6.5 · pp. 56–57
• 17–79% improvement in user quality of experience over baseline methods was reported for SR-Telemedicine’s adaptive video super-resolution.A scalable neural network and DDQN dynamically adjust resolution and model scale to network and computational capabilities.Source: §6.5 · p. 57
• TytoCare59 supports remote examinations using diverse medical data and computer vision and deep learning, while Proximie60 applies AR, vision, and deep learning to remote surgical collaboration and education.TytoCare enables real-time patient-provider consultations; Proximie supports procedural guidance and learning from recorded surgeries.Source: §6.5 · p. 57

7 Future Directions

• Future work spans privacy and security, energy efficiency and performance optimisation, adaptive architectures, multimodal fusion, efficient inference, explainability, quantum-inspired computing, and human-centric applications.These directions extend the survey’s discussion of edge deep learning in computer vision and medical applications.Source: §7 · p. 58

7.1 Privacy and Security

• Adversarial perturbations can mislead models and cause dangerous medical misdiagnoses, motivating adversarial training, regularisation, and resilient architectures.Adversarial training incorporates adversarial samples; cited work includes Isakov et al. 2019, Jha et al. 2021, and Vairo et al. 2023.Source: §7.1 · p. 58
• Homomorphic encryption, secure multiparty computation (SMPC), federated learning, and trusted execution environments can help protect sensitive data and computations on edge devices.Federated learning keeps patient information local while enabling collaboration across institutions; TEEs isolate sensitive computations.Source: §7.1 · p. 58

7.2 Energy Efficiency and Performance Optimisation

• Medical monitoring and real-time diagnostics in resource-constrained settings motivate edge-specific low-power processors and storage, including microcontrollers, ASICs, and neural network accelerators.Wearable ECG monitors and portable ultrasound scanners illustrate the need for immediate insights without cloud infrastructure; cited hardware work includes Li and Liewig 2020 and Reuther et al. 2021.Source: §7.2 · p. 58
• Data compression and lightweight neural networks can reduce computational and storage requirements while sustaining high performance.These are software-level approaches to energy efficiency and performance optimisation.Source: §7.2 · p. 59
• Energy-aware scheduling, adaptive voltage scaling, and power gating can reduce energy use while keeping wearable monitors ready to respond to critical events such as arrhythmias.Scheduling adapts task allocation and execution to device energy state and computational demands; low-power periods are particularly relevant to wearable health monitoring. Cited work includes Chang et al. 2019, Sarma and Biswas 2020, and Yousri et al. 2023.Source: §7.2 · p. 59

7.3 Adaptive Edge Computing Architectures

• Hardware-aware NAS considers device-specific latency, memory, and power constraints; ProxylessNAS searches on target hardware to balance latency and accuracy (Benmeziane et al. 2021; Cai et al. 2018).NAS can automatically optimize network design under resource constraints (Wistuba et al. 2019).Source: §7.3 · p. 59
• Slimmable Neural Networks dynamically adjust network width during inference, allowing one model to run at full or lighter capacity as resources vary (Yu et al. 2019).Source: §7.3 · p. 59
• Resource-aware dynamic pruning removes unimportant channels during inference based on input data, reducing complexity for less demanding tasks while optimizing power and performance.This enables models to adapt across different tasks, workloads, and device conditions.Source: §7.3 · p. 59

7.4 Multimodal Edge Computing Fusion Techniques

• Healthcare fusion combines medical imaging, electronic health records, and wearable-sensor data to support comprehensive patient assessments.Source: §7.4 · pp. 59–60
• Multimodal fusion must integrate diverse data while maintaining processing efficiency, generalisability, and accuracy for reliable, timely diagnoses.Source: §7.4 · p. 60
• ViLT jointly embeds images and text with minimal preprocessing for tasks such as medical report generation and cross-modal retrieval (Fields and Kennington 2023).Simplified cross-modal interactions may reduce computational overhead in edge environments.Source: §7.4 · p. 60
• Medical CLIP aligns radiology images with text reports through contrastive learning, supporting image retrieval and diagnostic support (Zhao et al. 2023).Contrastive learning maximises agreement between modalities while maintaining modality-specific representations.Source: §7.4 · p. 60

7.5 Efficient Edge Inference in Resource-Constrained Environments

• Edge inference efficiency depends on energy management and task scheduling across CPUs, GPUs, and FPGAs, with edge-cloud collaboration and real-time stream processing.Source: §7.5 · p. 60
• SparseNN skips computations involving zero-valued activations and weights, using network sparsity to reduce computational load.Source: §7.5 · p. 60
• ShiftAddNet replaces convolutional multiplications with shift-and-add operations, maintaining competitive performance while reducing latency and resource use (You et al. 2020).Source: §7.5 · p. 60

7.6 Explainable AI in Edge Models

• Transparency, trust, and regulatory compliance drive XAI in sensitive applications such as healthcare and autonomous driving.The section identifies a need for lightweight, real-time explainability frameworks suitable for edge devices.Source: §7.6 · p. 61
• LIME and SHAP provide visual explanations of how features influence model decisions.Source: §7.6 · p. 61
• Edge XAI still faces a trade-off between model performance and explainability, as well as a need for standardised explainability frameworks.Source: §7.6 · p. 61
• In healthcare, clinicians need understandable reasons for AI diagnoses and recommendations, and edge-based XAI can support clinical decisions and trust among patients and providers.For imaging analyses, explanations can make AI recommendations easier to justify and trust.Source: §7.6 · p. 61

7.7 Quantum-Inspired Edge Computing

• Quantum-inspired edge computing applies quantum-computing principles to edge methods, with potential to improve medical diagnostic data processing and resource allocation.Quantum ML is being explored as a way to enhance efficiency.Source: §7.7 · p. 61
• Quantum teleportation may enhance secure edge communications, while quantum annealing may address complex optimisation problems and improve performance.Source: §7.7 · p. 61
• Quantum-inspired edge technologies remain at an early stage, but may help overcome limitations of classical computing in edge environments.Source: §7.7 · p. 61
• Future work may develop more powerful quantum algorithms and quantum-resistant edge security protocols, and explore hybrid classical-quantum systems for diagnostic accuracy and speed.Source: §7.7 · p. 61

7.8 Human-Centric Edge Applications

• Portable edge devices can analyse health data instantly to deliver personalised feedback and advice, supporting early issue detection and proactive health management (Garc´ıa et al. 2017; Abrantes et al. 2023).These services help users better understand their health condition.Source: §7.8 · pp. 61–62
• Smart-home edge devices support user interaction and environmental monitoring, automating systems for comfort and energy efficiency; thermal imaging can guide room-temperature adjustment (Patchava et al. 2015; Schneider and Banerjee 2018; V´azquez and Kastner 2012).Source: §7.8 · p. 62
• Human-centric services face privacy, data-security, and computational constraints, making robust security protocols and careful handling of sensitive data important research challenges.Source: §7.8 · p. 62

7.9 Integration of Advanced Learning-Based Methods

• Transformers use self-attention to capture intricate patterns and long-range dependencies in medical images, while contrastive learning uses unlabeled data to learn robust representations and reduce reliance on large labeled datasets (You et al. 2022; You et al. 2022, 2023a,c).The passage associates both approaches with improved diagnostic accuracy or generalisation.Source: §7.9 · p. 62
• Few-shot learning addresses scarce medical labels with small training sets, while transfer learning fine-tunes or distills pretrained models for specific diagnostic tasks (You et al. 2022, 2023; You et al. 2021, 2022).Source: §7.9 · p. 62
• Domain adaptation aligns feature distributions across imaging conditions and equipment to improve generalisation, and generative models are among the surveyed diagnostic methods (You et al. 2020; You et al. 2018, 2023b).Source: §7.9 · p. 62
• Integrating advanced learning methods with edge computing can enhance medical diagnostic accuracy and efficiency, supporting more reliable clinical analysis and decision-making.Source: §7.9 · p. 62

8 Conclusion

• Edge DL brings deep learning closer to data sources for real-time processing and interpretation in resource-constrained settings.The survey focuses on computer vision and medical diagnostics and explores Edge DL’s fundamental concepts and technical merits.Source: §8 · p. 62
• Model compression and lightweight model design support Edge DL deployment on edge devices, with potential to reduce latency, conserve bandwidth, and bolster data privacy.Source: §8 · p. 62
• Continued hardware and software advances may expand Edge DL’s transformative impact across applications, while the survey highlights both its possibilities and challenges.The conclusion expresses hope that Edge DL can contribute to improved daily lives and universal healthcare accessibility.Source: §8 · pp. 62–63