Robust Lightweight Deep Learning Models for Oral Cancer Screening

Siddhant Bharadwaj; Aakash Shedsale; Tejashree Subramanya; Mohd. Azfar; Praveen Birur; Debnath Pal; Shankararama Sharma; Anupama Shetty; Rajesh Sundaresan

Source-linked AI summary

Robust Lightweight Deep Learning Models for Oral Cancer Screening

Siddhant Bharadwaj, Aakash Shedsale, Tejashree Subramanya, Mohd. Azfar, Praveen Birur, Debnath Pal, Shankararama Sharma, Anupama Shetty, Rajesh Sundaresan

arXiv:2608.21583v1Aug 21, 2026cs.AI

TL;DR

Smartphone oral-cancer screening lacks systematic evidence for lightweight models under real-world data and hardware constraints. This study optimises and stress-tests edge architectures, with the best model reaching 87.4% sensitivity and 97.2% negative predictive value against specialist labels.

Problem
Smartphone oral-lesion screening lacks systematic evaluation of lightweight edge-AI robustness and interpretability under class imbalance and acquisition artefacts.
Method
The study systematically evaluates and optimises convolutional, transformer, and hybrid architectures using approximately 30,000 multi-centre smartphone images and held-out testing.
Results
87.4% sensitivity and 97.2% Negative Predictive Value were achieved by the optimised edge system against specialist labels, outperforming primary healthcare-worker baselines.
Takeaways & Limitations
The findings support lightweight, interpretable edge AI as a potential approach for point-of-care oral-lesion triage in resource-constrained settings.
Takeaways & Limitations
External datasets were not used for validation, and single-specialist annotation prevented inter-annotator agreement and quantitative interpretability evaluation.

Abstract

from arXiv · show

Oral cancer is a leading cause of mortality in low-to-middle-income countries, where a shortage of specialists delays diagnosis. While point-of-care screening via smartphones offers a scalable solution, developing robust AI for resource-constrained settings poses significant challenges, including class imbalance in training data, variable data quality, and computational constraints on edge devices. In this paper, we present the optimisation of lightweight deep learning models for smartphone-based oral cancer screening. Using a diverse, multi-centre retrospective dataset of approximately 30,000 images acquired over a decade, we systematically evaluate state-of-the-art convolutional, transformer, and hybrid architectures. Through rigorous pipeline ablation, we demonstrate that directly optimising hybrid architectures for the edge strictly outperforms computationally heavy paradigms, such as large models or knowledge distillation. Furthermore, interpretability analysis and simulated noise-stress tests revealed that the system anchors on clinical features and remains robust to unstructured sensor noise, despite vulnerabilities to impulse bit errors. In the held-out test set, our optimised MobileViTv2 models achieved an average sensitivity of 83.2 $\pm$ 1.5% and an average specificity of 86.0 $\pm$ 0.8%, with the best model exhibiting 87.4% sensitivity, 86.5% specificity, and a critical negative predictive value of 97.2% with reference to specialist labels. These results confirm that with targeted architectural selection and streamlined optimisation, interpretable and robust lightweight AI models exhibit high potential for edge deployment to enable automated triage in primary care settings.

I. INTRODUCTION

The introduction frames oral cancer screening as an urgent low-resource diagnostic challenge and motivates smartphone-based deep learning as a point-of-care solution. The study develops and evaluates lightweight models using a diverse, multi-centre dataset while addressing edge deployment, interpretability, and robustness.

Clinical motivation: Oral cancer causes substantial mortality in low-to-middle-income countries, while limited specialist access contributes to late detection and variable primary-care sensitivity.India alone records over 140,000 new cases annually, and field studies report sensitivities as low as 43.6% compared to specialists.
Clinical motivation: Smartphone imaging with deep learning is feasible for point-of-care lesion classification, but existing studies rely largely on small curated datasets without systematic real-world evaluation.Primary-care data are affected by class imbalance and physical acquisition artefacts.
Study contribution: The study uses approximately 30,000 smartphone-captured images from a diverse, multi-centre retrospective dataset collected over a decade to develop robust lightweight screening systems.The dataset is used to examine technical challenges and build a lightweight model.
Study contribution: MobileViTv2 is identified as the optimal inference engine among evaluated convolutional, transformer, and hybrid architectures, combining transformer contextual awareness with convolutional feature extraction.The paper also reports that direct optimisation of lightweight models on ground-truth labels outperforms compute-heavy large-model paradigms, while knowledge distillation adds overhead without significant improvement.
Study contribution: Interpretability analyses associate predictions primarily with annotated clinical features, while hardware stress tests assess low-light sensor noise and impulse bit-errors.The framework reports 87.4% sensitivity and 97.2% Negative Predictive Value with on-device inference capabilities.

II. MATERIALS AND METHODS … 2) Data Description:

The study used a retrospective, multi-centre dataset of oral images collected in rural Indian outreach programmes and rigorously cleaned, labelled, and partitioned for model development and unbiased evaluation.

A. Retrospective Data: A retrospective dataset supported AI model development, tuning, and evaluation through training, validation, and strictly held-out test sets.The validation set was used for checkpointing, while the held-out test set provided final performance benchmarking.
1) Data Collection:: Images came from Biocon Foundation outreach programmes across northern, north-eastern, and southern India between 2011 and 2023, primarily targeting rural populations.
1) Data Collection:: Frontline health workers captured images using heterogeneous consumer-grade smartphones, including HTC, Motorola, Xiaomi, and Samsung devices.This collection design introduced device heterogeneity relevant to systems deployment.
2) Data Description:: 31,601 raw images were annotated by two oral medicine specialists before quality-control filtering.
2) Data Description:: 29,574 images remained after removing 126 statistically identical duplicates and non-diagnostic images such as extreme blur, out-of-focus images, and insufficient mouth opening.
2) Data Description:: Each cleaned image received a provisional clinical diagnosis that was subsequently mapped into a binary Suspicious versus Non-suspicious classification schema.
2) Data Description:: The dataset split demonstrated real-world class imbalance across the training, validation, and test partitions.The supplied passage identifies this datasplit as illustrating real-world class imbalance but does not provide the partition counts.

3) Datasplit and Class Imbalance: … 2) Data Augmentations for Real-World Robustness:

The study uses a 60-20-20 data split and addresses severe class imbalance while evaluating lightweight architectures and robustness-oriented training procedures for smartphone-based screening. Training combines standardized implementation choices with conventional augmentations and exploratory SAM-based removal of extraneous regions.

3) Datasplit and Class Imbalance:: 60-20-20 split allocated images to training, validation, and testing, while suspicious-to-non-suspicious images occurred at approximately a 1:5 ratio.The imbalance necessitated targeted optimisation strategies during training.
B. Lightweight AI Architectures: 7.5 million parameters was the upper limit for architectures selected to remain computationally viable on budget smartphones.The search covered six architectures across CNN, pure Vision Transformer, and hybrid CNN-Transformer families.
B. Lightweight AI Architectures: Six lightweight architectures were evaluated: EfficientNetV2-B0, MobileNetV3-Large, NASNet-Mobile, DeiT-Ti, MobileViTv2, and EdgeNeXt-S.The candidates spanned convolutional, pure transformer, and hybrid CNN-Transformer designs.
1) Training Infrastructure and Hyperparameters:: 50 epochs with a batch size of 32 were used for training, and the checkpoint with the best validation performance was evaluated on the test set.Training used PyTorch 2.3 on 3 NVIDIA RTX 3090 GPUs, with ImageNet-pre-trained initialization.
2) Data Augmentations for Real-World Robustness:: Zoom-in, colour jitter, horizontal/vertical flips, and random rotations between -10° and 10° simulated variability in smartphone-captured images.These conventional geometric and photometric augmentations were also intended to prevent over-parameterisation.
2) Data Augmentations for Real-World Robustness:: SAM “everything” mode generated coarse masks that removed non-diagnostic regions to test whether models would focus on the oral cavity rather than background artefacts.SAM was explored as an augmentation because accurate bounding boxes were unavailable for all images.

3) Image Quality Reweighting: · 4) Handling Systemic Class Imbalance:

The study tested perceptual image-quality reweighting using NIQE and CLIP-IQA, alongside five strategies to address the dataset’s systemic imbalance between suspicious and non-suspicious images. These methods respectively aimed to emphasize higher-quality inputs and prevent majority-class prediction.

3) Image Quality Reweighting:: NIQE and CLIP-IQA were evaluated as no-reference image-quality metrics for reweighting the training loss.NIQE measures deviations from natural scene statistics, whereas CLIP-IQA evaluates perceptual quality through semantic consistency with natural image priors.
3) Image Quality Reweighting:: The inverse NIQE score was integrated into the loss to dynamically down-weight low-quality images.Lower NIQE values indicate higher image quality.
3) Image Quality Reweighting:: CLIP-IQA scores were used as weighting factors to test whether high-level perceptual understanding could guide training.CLIP-IQA assesses perceptual quality based on semantic consistency with natural image priors.
4) Handling Systemic Class Imbalance:: 2,932 suspicious and 14,812 non-suspicious images produced a five-to-one training imbalance requiring mitigation.The imbalance was described as systemic and typical of screening environments.
4) Handling Systemic Class Imbalance:: Random undersampling matched the majority non-suspicious class to the minority suspicious-class count.The majority class was randomly undersampled.
4) Handling Systemic Class Imbalance:: Random oversampling stochastically increased the number of minority suspicious-class samples.This strategy was evaluated among five imbalance-mitigation methods.
4) Handling Systemic Class Imbalance:: Weighted Cross-Entropy assigned greater loss penalties to minority-class errors using inverse class-frequency weighting.The training-split weights were w0 = 0.60 for non-suspicious and w1 = 3.03 for suspicious images.
4) Handling Systemic Class Imbalance:: Focal Loss emphasized hard examples by down-weighting well-classified background examples, while Balanced Mix-up combined instance-based and class-balanced sampling.Balanced Mix-up used λ ∼Beta(α, 1), with α set to 0.1, 0.2, and 0.3.

III. EXPERIMENTAL RESULTS · A. Data Augmentation and Pipeline Efficiency

The experiments identified geometric transformations as the most consistently effective augmentation strategy across lightweight architectures, while photometric, zoom-in, and segmentation-based methods introduced limitations. Excluding SAM streamlined data curation without compromising diagnostic accuracy.

III. EXPERIMENTAL RESULTS: “Best Pair” denotes the sensitivity and specificity of the top-performing model by geometric mean among 10 fold-wise checkpointed models.This metric is based on fold-wise checkpointed models.
A. Data Augmentation and Pipeline Efficiency: Four augmentation strategies were systematically evaluated: segmentation-based, zoom-in, geometric transformations, and colour jitter.The comparison covered all four strategies across lightweight architectures.
A. Data Augmentation and Pipeline Efficiency: Geometric transformations, specifically horizontal-vertical flips with random rotations, produced consistent sensitivity and specificity improvements across tested lightweight architectures.They emerged as the most effective augmentation strategy in the evaluation.
A. Data Augmentation and Pipeline Efficiency: Colour jitter degraded metrics for NASNet-Mobile, EfficientNetV2-B0, and MobileNetV3-Large.Aggressive chromatic modification may obscure diagnostic colour distinctions such as red versus white lesions.
A. Data Augmentation and Pipeline Efficiency: Zoom-in augmentation had inconsistent effects, with loss of contextual texture hampering the receptive fields of certain architectures.Its impact varied across models rather than yielding a uniform improvement.
A. Data Augmentation and Pipeline Efficiency: SAM-based segmentation augmentation frequently generated erroneous masks because precise bounding boxes were unavailable.Although segmentation guidance was intended to improve focus, qualitative review exposed unreliable masks.
A. Data Augmentation and Pipeline Efficiency: Excluding SAM removed a computationally intensive, high-memory dependency and streamlined training without compromising diagnostic accuracy.The benefit applied to the data curation and training pipeline from a systems-engineering perspective.

B. Reweighting based on Image Quality · C. Strategies for Handling Class Imbalance

Image-quality reweighting offered no significant overall advantage and added computational overhead, while class-imbalance experiments favored Weighted Cross-Entropy for improving sensitivity without compromising efficiency and reliability.

B. Reweighting based on Image Quality: Image-quality reweighting produced performance largely comparable to baseline, with no significant overall advantage.Trade-offs varied by architecture, and EdgeNeXt-S, DeiT-Ti, and NasNet-Mobile showed reduced average sensitivity without compensatory specificity gains.
B. Reweighting based on Image Quality: Generic perceptual image-quality metrics may misalign with diagnostic utility by down-weighting noisy images that retain discriminative clinical features.Field-collected training images are expected to contain noise, potentially causing many useful examples to receive lower weights.
B. Reweighting based on Image Quality: Complex perceptual image-quality models added unnecessary training-pipeline overhead without significant downstream performance gains.
C. Strategies for Handling Class Imbalance: Random undersampling delivered the largest sensitivity increase among sampling strategies but slightly reduced specificity.Random oversampling produced a more conservative improvement while increasing training burden by inflating dataset size.
C. Strategies for Handling Class Imbalance: Weighted Cross-Entropy consistently improved sensitivity across all models while adding no training overhead and preserving the original data distribution.Its loss weights were based on inverse class frequency, unlike oversampling, undersampling, or Mix-up.
C. Strategies for Handling Class Imbalance: Focal Loss caused architectural instability, including performance inversions in MobileNetV3-L, EfficientNetV2-B0, and NASNet-Mobile.
C. Strategies for Handling Class Imbalance: Balanced Mix-up increased sensitivity but sharply reduced specificity, producing significantly more false positives for triage.The resulting false-positive increase was undesirable for a triaging system.
C. Strategies for Handling Class Imbalance: Weighted Cross-Entropy provided the strongest balance between minority-class detection, overall system reliability, and computational efficiency.

D. Final Model Architecture Selection and Five-fold Cross Validation · E. Knowledge Distillation Ablation

Geometric augmentations with Weighted Cross-Entropy formed the most efficient training strategy, while MobileViTv2 was the most robust architecture and showed stable five-fold performance. Knowledge distillation introduced instability, decision-boundary bias, and computational overhead without robust gains, so the final framework used direct lightweight-model training.

D. Final Model Architecture Selection and Five-fold Cross Validation: Geometric augmentations combined with Weighted Cross-Entropy were identified as the most efficient training strategy.
D. Final Model Architecture Selection and Five-fold Cross Validation: MobileViTv2 emerged as the most robust candidate, whereas EfficientNetV2-B0, MobileNetV3-Large, NASNet-Mobile, DeiT-Ti, and EdgeNeXt-S were less consistent or lower-performing.
D. Final Model Architecture Selection and Five-fold Cross Validation: MobileViTv2 combines convolutional local-feature extraction with transformer global-context modelling in a constrained ∼4.9M parameter footprint.
D. Final Model Architecture Selection and Five-fold Cross Validation: 83.2 ± 1.5% mean sensitivity and 86.0 ± 0.8% mean specificity confirmed systemic stability across five MobileViTv2 cross-validation folds.The low standard deviations indicated robust generalisation and reduced concern about fold-specific overfitting.
E. Knowledge Distillation Ablation: Knowledge distillation was excluded because it added pipeline complexity, convergence instability, and capacity bottlenecks without robust performance improvements.The final deployment framework instead directly trained lightweight models on ground-truth labels.
E. Knowledge Distillation Ablation: 100% EfficientNetV2-L distillation caused model collapse, reducing average sensitivity to 38.16 ± 32.92%.The paper attributes this mainly to architectural inductive-bias mismatch and partly to capacity mismatch.
E. Knowledge Distillation Ablation: 100% Swin-B distillation raised sensitivity to 86.43 ± 1.85% but reduced specificity to 62.65 ± 2.63%.The 50% configuration restored balance at 83.50% sensitivity and 87.37% specificity, supporting blended hard- and soft-label supervision.
E. Knowledge Distillation Ablation: 84.22 ± 3.56% sensitivity and 87.37 ± 3.53% specificity from 50% ViT-B/16 distillation remained insufficient to justify teacher-model compute and memory overhead.

IV. INTERPRETABILITY VALIDATION

Interpretability validation used GradCAM++ and Attention Rollout to assess whether MobileViTv2 relied on clinically relevant signals. Activations aligned with specialist lesion annotations, while Transformer attention captured multi-scale oral-cavity context.

Interpretability methods: GradCAM++ evaluated local convolutional features, while Attention Rollout mapped global Transformer context across spatial scales.The dual-technique analysis matched MobileViTv2’s hybrid architecture and hierarchical design.
Local feature attribution: GradCAM++ heatmaps showed tight spatial alignment with ground-truth segmentations provided by clinical specialists.The model’s highest activations closely aligned with expert lesion annotations, indicating avoidance of spurious background correlations.
Global contextual attribution: Transformer attention maps confirmed that MobileViTv2 aggregated multi-scale contextual features from the oral cavity to inform inference.Attention Rollout demonstrated global-context extraction at feature scales 256, 64, and 16.

V. ROBUSTNESS AND STRESS TESTING

Stress tests found that the optimised MobileViTv2 model remained viable under distributed Gaussian noise but degraded substantially under structured stripe and impulse noise. These failures were linked to lesion masking and the loss of structural features during preprocessing.

Stress-test design: Synthetic noise perturbations were applied at varying intensities to raw held-out test images to evaluate hardware-simulated acquisition failures.The tests covered unstructured Gaussian, structured stripe, and impulse salt-and-pepper noise.
Gaussian noise: Over 68% sensitivity persisted at 20% Gaussian noise, indicating robustness to distributed, low-level sensor noise.Gaussian noise simulated low-light sensor noise and produced predictable, monotonic degradation.
Stripe noise: At 2% stripe noise, performance matched 15% Gaussian noise, while 7% stripe noise caused a significant sensitivity drop.Structured stripes can mask lesion areas and lead to missed detections.
Impulse noise: 5.1% sensitivity and 99.3% specificity occurred at 7% impulse noise, revealing severe vulnerability to salt-and-pepper corruption.High-frequency disturbances propagated during resizing, obliterating edges, textures, and other distinctive diagnostic features.

VI. DISCUSSION · A. Architectural Findings and Pipeline Efficiency · B. Potential for Clinical Efficacy at the Edge

The discussion identifies MobileViTv2 hybrid architectures and direct ground-truth optimisation as efficient strategies for edge deployment. It also reports specialist-benchmarked diagnostic performance, high negative predictive value, and potential to expand oral cancer triage in resource-constrained settings.

VI. DISCUSSION: Heterogeneous smartphone cameras and unconstrained lighting increased data-curation challenges but enhanced final-system generalisability by encouraging robust feature learning.
A. Architectural Findings and Pipeline Efficiency: MobileViTv2 offers the best balance of computational efficiency and diagnostic performance among evaluated architectures.Its separable self-attention computes global context with linear O(N) complexity rather than the quadratic O(N^2) complexity of standard Vision Transformers.
A. Architectural Findings and Pipeline Efficiency: Zero-shot Segment Anything Model masks frequently occluded relevant mucosal boundaries, introducing noise instead of clarity.
A. Architectural Findings and Pipeline Efficiency: Knowledge Distillation from ViT and EfficientNetV2-L teachers created severe training overhead without commensurate performance gains.Complete distillation also produced catastrophic model collapse, indicating that ultra-lightweight models lack sufficient parameter space to converge reliably on complex soft-label distributions.
A. Architectural Findings and Pipeline Efficiency: Directly optimising lightweight architectures with ground-truth labels is advocated over mimicking bloated, compute-intensive models.Future systems-level work could replace heavy generalised foundation models with lightweight architectures trained for domain-specific tasks.
B. Potential for Clinical Efficacy at the Edge: 87.4% sensitivity and 86.5% specificity were achieved by the optimised deep-learning system on held-out test images against specialist benchmarks.Reported field-worker sensitivity is 43.6%-60%, with slightly higher specificities of 78%-81%.
B. Potential for Clinical Efficacy at the Edge: 17 MB model size enables deployment in resource-constrained healthcare networks and supports the potential to democratise oral cancer screening.
B. Potential for Clinical Efficacy at the Edge: 97.2% Negative Predictive Value was achieved by the best-performing checkpoint, supporting effective exclusion of non-suspicious cases in triage.The discussion identifies false negatives as a fatal referral-delay risk and false positives as a burden on secondary-care facilities.

C. Limitations and Future Directions · VII. CONCLUSION

The study establishes a lightweight edge framework for oral-lesion triage with performance comparable to large models at lower compute and memory requirements, while identifying limitations in dataset scope, annotation, noise robustness, and longitudinal adaptation. Specialist-benchmarked results support its potential for AI-assisted screening in resource-limited settings, alongside future needs for risk stratification and multimodal patient-level analysis.

C. Limitations and Future Directions: The development dataset came from India alone, although screening programmes spanned a decade and a wide geographic area.
C. Limitations and Future Directions: Future work could combine image tensors with patient metadata, develop granular multi-class classification, and evaluate continual learning for hardware and demographic drift.
C. Limitations and Future Directions: Only one specialist reviewed each image, preventing inter-annotator agreement reporting and potentially introducing subjective interpretations and biases.
C. Limitations and Future Directions: Specialist-segmented lesion masks were available for only 244 images, preventing derivation of a quantitative interpretability performance measure from the described mechanisms.
C. Limitations and Future Directions: The model showed vulnerabilities to stripe and salt-and-pepper noise, motivating frequency filtering, destriping, denoising, and related methods to prevent predictable failures.
VII. CONCLUSION: Directly optimising lightweight architectures achieved performance comparable to large models with significantly reduced compute and memory requirements.
VII. CONCLUSION: 87.4% specialist-benchmarked sensitivity and 97.2% Negative Predictive Value significantly outperformed baselines of primary healthcare workers.
VII. CONCLUSION: The framework establishes a robust foundation for future edge deployments and could enable AI-assisted oral cancer screening in resource-limited settings.