PaperlayerView exact revision on arXiv

Source-linked AI summary

LukeNet: A lightweight CNN integrated with an XAI model for Smart acute lymphoblastic leukemia detection and management

Md Taimur Ahad

arXiv:2609.31736v1cs.CVcs.AIcs.NE

TL;DR

Clinical adoption of CNN-based ALL detection remains challenging despite laboratory performance, and edge devices constrain memory and computation. The paper presents LukeNet, a CNN applied to three datasets with SHAP and Grad-CAM explainability. It reports 99% accuracy on unseen test data.

  • Problem

    Clinical adoption of CNNs remains limited by deployment and hospital-integration challenges, while edge devices have limited memory and computational capacity.

    Source: §1 · pp. 4–5

  • Method

    LukeNet is a CNN requiring minimal preprocessing, applied to three datasets, with SHAP and Grad-CAM used to explain predictions.

    Source: §1 · p. 6, §6 · p. 42

  • Results

    LukeNet achieved 99% accuracy on unseen test data.

    Source: §6 · p. 42

  • Takeaways & Limitations

    SHAP and Grad-CAM visualizations highlighted pixel-level feature contributions and influential image regions, improving interpretability and clinical trust in automated predictions.

    Source: §1 · p. 6

    Inference ran at 1 FPS, which the authors characterize as near-real-time rather than real-time; further optimization is needed for clinical deployment.

    Source: acknowledge that the inference speed… · p. 42

Abstract

from arXiv · show

Acute Lymphoblastic Leukemia (ALL) patients require early, accurate detection to enable timely treatment and effective patient management. A Convolutional Neural Network (CNN) is well-suited for creating an end-to-end enabling environment for ALL detection and classification. However, most CNN-based ALL detection systems are theoretical and unsuitable for deployment on edge devices due to high computational demands. The Internet of Medical Things (IoMT)-enabled devices offer an opportunity to monitor ALL patients in real time. Wearables that track temperature, heart rate, oxygen saturation, and activity can deliver critical data to support timely clinical intervention and improve patient outcomes. In smart IoMT environments, a lightweight CNN is essential because connected devices often operate under limited computational power, memory, and latency constraints. To address this need, this study proposes LukeNet, a lightweight CNN integrated with explainable artificial intelligence (XAI) for an IoMT-based SMART Acute Lymphoblastic Leukemia Detection and Management System. Trained on three (3) ALL datasets and five-fold cross-validation, LukeNet achieved an impressive 99% model accuracy as well as 99% unseen test accuracy, which is higher than six state-of-the-art (SOTA) CNNs, such as DenseNet121, MobileNet, ResNet50, InceptionV3, Xception, and VGG16, as well as transfer learning models. Furthermore, LukeNet was compared with two ensemble models. In addition, explainable AI methods are integrated to highlight relevant regions in microscopic images. The novelty of this study lies in the architecture of LukeNet, which balances model depth and computational efficiency by using depthwise separable convolutions, mitigates the risk of gradient loss in deeper layers, and provides strong global and local feature extraction capabilities.

1. Introduction

The introduction motivates a lightweight, interpretable CNN for ALL detection in smart healthcare settings, where edge-device constraints and clinical translation remain challenges. It presents LukeNet and its evaluation and explainability contributions.

1. Introduction

  • ALL's rapid malignant-cell expansion and potentially severe organ involvement frame the need for early diagnosis and structured clinical management.The introduction describes PBS as an initial screening test and notes that human interpretation can be labor-intensive and error-prone.Source: §1 · pp. 3–4
  • Edge devices' limited memory and computational capacity motivate lightweight CNN inference for low-latency, low-power diagnostic support in smart leukemia management.The introduction describes smart systems as supporting continuous, real-time vital-sign and tumor-marker tracking, remote healthcare, and personalized treatment.Source: §1 · pp. 4–5
  • The study identifies limited preprocessing alongside high accuracy as a CNN design gap, with modality variation, feature-map correlations, poor generalization, and gradient loss among the cited concerns.The introduction also notes concerns about low accuracy and false-positive rates in CNN-based ALL detection.Source: §1 · p. 5
  • LukeNet is presented as a CNN requiring minimal image processing, with its robustness evaluated on three microscopic PBS datasets.The introduction frames its design as addressing the risk of gradient loss as network depth increases.Source: §1 · pp. 5–6
  • LukeNet was applied to three microscopic PBS datasets and evaluated against six state-of-the-art CNNs trained from scratch and with transfer learning.The six models were DenseNet121, ResNet50, InceptionV3, Xception, MobileNet, and VGG16.Source: §1 · p. 6
  • SHAP highlighted pixel-level feature contributions, while Grad-CAM localized image regions influential in LukeNet's decisions.The introduction states these tools improved interpretability and clinical trust in automated predictions.Source: §1 · p. 6

2. State of the Art

Prior ALL-detection studies used customized CNNs, CNN-XAI combinations, and multistage deep-learning pipelines, reporting varied classification results across datasets and tasks.

2. State of the Art

  • Anand et al. (2025) used a customized five-block CNN and reported accuracy of 0.96 and precision of 0.95 with the Adam optimizer.The described blocks included 13 convolutional layers and five max-pooling layers.Source: §2 · p. 6
  • Thiriveedhi et al. (2025) integrated CNNs and XAI for ALL diagnosis and interpretation, reporting average precision of 99.35%, recall of 99.33%, and F1 score of 99.58%.The passage says the metrics exceeded several listed models except DenseNet201.Source: §2 · p. 7
  • Other cited approaches reported 99.61% accuracy for DeepLeukNet, 97.18% for Claro (2020), and 98.61% for LeukNet cross-validation.LeukNet's cross-dataset accuracies were 97.04%, 82.46%, and 70.24%.Source: §2 · p. 7
  • Vo et al. (2022) combined modified-UNet segmentation, CNN classification, and perceptron data fusion, reporting 93% overall accuracy, 95% ALL detection, and 92% malaria detection.Sampathila (2022) reported ALLNET accuracy of 95.54%, specificity of 95.81%, sensitivity of 95.91%, F1-score of 95.43%, and precision of 96%.Source: §2 · p. 7

3. Research Methodology

The methodology evaluates LukeNet as a lightweight ALL classifier across three datasets, using stratified five-fold validation and comparisons with other CNN, transfer-learning, and ensemble models. It describes the dataset splits, image augmentation, model architecture, training pipeline, and hyperparameter tuning.

3. Research Methodology

  • LukeNet was designed as a lightweight, edge-integratable ALL PBS classifier, with high accuracy on unseen datasets as a design criterion.It was tested on three ALL datasets.Source: §3 · p. 8

3.1 Experiment Platform and Dataset

  • Dataset A has four classes and 3,242 images; Dataset B has five classes with 1,000 images per class; Dataset C has eight classes and 17,092 ALL images.The datasets were split 70% for training, 20% for validation, and 10% for testing.Source: §3.1 · pp. 10–11

3.2 Image augmentation process

  • Training images were augmented through scaling, cropping, flipping, rotation, distortion, shear, skew, and intensity or color transformations.Random rotations included angles from -15 to 15 degrees, as well as rotations at multiples of 90 degrees.Source: §3.2 · pp. 11–12

3.3 LukeNet model description

  • LukeNet uses six convolutional layers, one depthwise-separable layer, global average pooling, and regularized dense layers with dropout.Filters increase from 32 to 1,024; depthwise separation reduces parameters, while batch normalization, GAP, L1/L2 regularization, and dropout are also used.Source: §3.3 · pp. 12–14

3.4 LukeNet training process

  • Training passes images through convolution and normalization, a depthwise-separable block, GAP, regularized dense layers, and a SoftMax class-probability output.Source: §3.4 · pp. 16–18

3.5 Training hypermeters

4. Results of experiments

LukeNet’s reported accuracy was 98%, 100%, and 98% across three datasets; benchmark comparisons describe strong performance while also noting strong DenseNet121 and Xception results. The DVS ensemble’s curves showed no overfitting.

4. Results of experiments

  • 979 Early, 928 Pre, and 794 Pro instances were correctly classified; the 8-class dataset included 46 Ig, 20 Lymphocyte, and 57 Monocyte misclassifications.The confusion-matrix description also reports relatively few Type 1 errors in the 5-class and 8-class datasets.Source: Figure 8 · p. 19, §4 · p. 20

4.1 Classification Report

  • 98%, 100%, and 98% accuracy were reported for Dataset-A, Dataset-B, and Dataset-C, respectively, with some Basophil and Ig misclassifications in Dataset-C.Precision, recall, and F1-scores were high across most Dataset-C classes.Source: Table 3 · p. 20, §4.1 · p. 21

4.2 Training accuracy and Data loss curve

  • Training and validation accuracy and loss diverged somewhat for Dataset-A, with validation metrics lagging training metrics and indicating some overfitting.Training accuracy rose rapidly and training loss fell sharply, while validation accuracy increased more gradually and validation loss remained slightly higher.Source: §4.2 · p. 21, Table 4 · p. 21, Figure 9 · p. 22

4.3 Benchmark comparison

  • LukeNet was described as more reliable and accurate, while DenseNet121 and Xception performed consistently well and transfer learning generally helped Xception and MobileNet.VGG16 showed no significant improvement with transfer learning.Source: §4.3 · pp. 24–25, Table 5 · p. 24

4.4 Ensemble model

  • The DVS ensemble combines DenseNet201, VGG19, and SE-ResNet152, and its accuracy and loss curves were reported to show no overfitting.The cited curves compare training and validation over epochs.Source: §4.4 · pp. 25–26, Figure 11 · p. 26

4.5 Ensemble model-2

The DIX ensemble achieved 99.12% accuracy and was reported to outperform the original CNN and transfer learning. LukeNet was also reported to use fewer resources across several computational measures.

4.5 Ensemble model-2

  • 99.12% accuracy was reported for DIX, with 0.83% better performance than the original CNN architecture and 4.12% better performance than transfer learning.DIX combines DenseNet201, InceptionV3, and Xception.Source: §4.5 · pp. 26–27, Table 8 · p. 26

4.6 Computational cost of the LukeNet and deep learning models

  • LukeNet’s parameter count remains consistent across classification tasks, and its memory usage is reported as the lowest among SOTA models.The comparison covers parameter count and memory usage.Source: §4.6 · p. 28, Figure 10 · p. 29
  • LukeNet was reported to have significantly lower latency in milliseconds per image than many SOTA models, including MobileNet on more complex tasks.The comparison is shown as latency in ms/image.Source: §4.6 · p. 28, Figure 10 · p. 29
  • LukeNet requires far fewer floating-point operations than models such as MobileNet, according to the GFLOPS comparison.The passage describes LukeNet as computationally efficient.Source: §4.6 · p. 28, Figure 10 · p. 30
  • LukeNet’s inference memory consumption was reported as significantly lower than DenseNet121 and VGG16.The comparison is shown in the inference-memory panel.Source: §4.6 · p. 28, Figure 10 · p. 30
  • LukeNet’s training memory usage was described as minimal relative to other SOTA networks and consistent across task complexity.The passage also characterizes its training-memory footprint as low.Source: §4.6 · p. 28, Figure 10 · p. 31

4.1 Descriptive statistics analysis

LukeNet’s descriptive statistics report a 91% mean accuracy and a 15 standard deviation. The section also describes pixel- and region-based explanations of model predictions.

4.1 Descriptive statistics analysis

  • 91% mean accuracy, a standard deviation of 15, and an 85%–97% confidence interval were reported for LukeNet.The passage describes the mean as close to ResNet50 and Inception V3.Source: §4.1 · p. 31, Table 8 · p. 31, Table 9 · p. 32

4.7 Explainability with SHAP and Grad-CAM

  • SHAP visualizations attribute predictions to pixels, while LIME visualizations highlight diagnostically relevant structures such as dense masses or irregular textures.The passage describes SHAP heatmaps as showing pixel contributions and LIME views as illustrating spatial attention.Source: §4.7 · p. 32, Figure 11 · p. 33, Figure 12 · p. 34
  • Grad-CAM heatmaps highlighted cancerous areas in correctly classified positive images and focused on non-cancerous features in negative images.The passage presents these views as spatial attention during classification.Source: §4.7 · p. 32, Figure 13 · p. 35, Figure 14 · p. 36

5. Concept of Smart acute lymphoblastic leukemia detection and management

The proposed system places the trained LukeNet model on a machine-learning server connected to imaging inputs and user devices. The application was tested on a Realme Note 70, where it processed one frame per second.

System architecture

  • The proposed system connects an IoT camera, gateway, machine-learning server, and router, with the trained model residing on the server.Edge devices act as delivery nodes, and the system proposes wireless connections.Source: §5 · p. 37, Figure 26 · p. 38

System workflow

  • Histopathology images can enter through hospital imaging systems or portable devices, and classification results with probabilities are displayed on web or mobile interfaces.The described interface shows the leukemia class label and confidence score alongside the test image.Source: §5 · pp. 37–38, Figure 27 · p. 39, Figure 28 · p. 41

Continuous monitoring

  • The proposed monitoring layer uses IoT devices to track symptoms and transmit monitoring data to healthcare providers.Examples include wearable health monitors and remote patient monitoring devices.Source: §5 · p. 41

Edge-device test

  • On a Realme Note 70, LukeNet processed 1 FPS and battery use was 21%; CPU temperature was recorded as N/A, with no significant increase reported.The passage describes the device test as an edge-device application run.Source: §5 · pp. 41–42

6. Discussions

The discussion reports LukeNet’s unseen-test performance alongside comparisons of original CNNs, transfer learning, and weighted ensembles. It also notes that several cited studies achieved higher accuracy than the reported ensemble.

6. Discussions

  • 99% unseen test accuracy was reported for LukeNet, with better accuracy than six state-of-the-art CNNs.The compared models were DenseNet121, ResNet50, InceptionV3, Xception, MobileNet, and VGG16.Source: §6 · p. 42
  • 98% accuracy made DenseNet201 the best-performing original CNN in the study.The discussion relates DenseNet’s connectivity pattern to information flow through the network, while noting challenges from gradient-based loss functions and class imbalance.Source: §6 · p. 43
  • Transfer learning yielded slightly higher accuracy than original individual networks on datasets with fewer than 2,000 images.SE-ResNeXt101 improved by 17% after transfer learning; ImageNet- or MURA-trained images improved results relative to no transfer learning.Source: §6 · p. 43
  • 98.76% accuracy was achieved by the weighted ensemble, although four cited studies reported better results.The discussion also says an ensemble improved accuracy compared with a single CNN architecture.Source: §6 · pp. 43–44

7. Contributions

The contributions center on LukeNet’s minimal image processing and high accuracy, experiments across datasets and architectures, and a weighted ensemble. They also report findings about architecture performance and negative transfer.

7. Contributions

  • LukeNet requires minimal preprocessing and achieved 99% to 100% accuracy across ALL detection tasks across three datasets.The study presents experiments across datasets and methods rather than relying on a single model or method.Source: §7 · p. 44
  • The contribution list positions LukeNet as a high-accuracy Deep CNN requiring minimal image processing for clinical settings.Source: §7 · p. 44
  • DenseNet201 achieved 98.08% accuracy, compared with 96.99% for ResNet152v2, 96.94% for VGG19, 90.93% for SE-ResNet152, and 86.41% for ResNext101.The study groups these models as multi-path, depth-based, spatial exploitation, feature-map exploitation, and width-based multi-connection architectures, respectively.Source: §7 · p. 44
  • The study reports that negative transfer is caused by dissimilar target and source datasets.Source: §7 · p. 45
  • The proposed DVS ensemble combines DenseNet201, SE-ResNet152, and VGG19 and achieved 98.76% accuracy.Saving individual CNN weights in .h5 files for testing took less time than other ensembling approaches.Source: §7 · p. 45

8. Limitations and Future Scopes

The paper identifies uncertainty about generalization beyond the dataset used for testing and outlines broader data and deployment work as future directions.

8. Limitations and Future Scopes

  • The authors state that testing on only one dataset leaves performance on other datasets uncertain.Source: §8 · p. 45
  • The approach was tested on PBS images; future work could tailor it to mammography, histopathology, and MRA images.Source: §8 · p. 45
  • The study has not yet been applied, and the authors propose future deployment in a machine applicable to CAD.Source: §8 · p. 45

9. Conclusion

The conclusion presents LukeNet as a compact, explainable CNN for PBS classification, while noting imaging and deployment directions that remain. The paper also provides data-availability and funding statements.

9. Conclusion

  • LukeNet has 11 million parameters and achieved over 99% accuracy across three datasets, surpassing SOTA CNNs, transfer-learning models, and ensemble models.The conclusion describes SHAP and Grad-CAM as post-hoc explainability tools and says the model is currently limited to grayscale mammograms.Source: §9 · pp. 45–46
  • The conclusion identifies real-time edge deployment, clinical validation, and collaboration with radiologists as future work.It also proposes exploring multi-view or temporal mammography and integration with other modalities.Source: §9 · p. 46

Data availability statement

Funding statement

Credit authorship contribution statement

The section assigns authorship contributions to Md Taimur Ahad and declares no known competing interests, while thanking anonymous reviewers.

Credit authorship contribution statement

Declaration of Competing Interest

Loading 2609.31736v1…