Lightweight Medical Imaging: Lessons from HI-MobileNet
· Anandhu P Shaji · IEEE RAICS 2026 Paper
In medical machine learning, papers routinely claim state-of-the-art accuracy by scaling to multi-billion-parameter Vision Transformers (ViTs) and ensemble backbones. In academic benchmarks with unmetered cluster GPUs, this works. In community health clinics across rural and underserved regions, the reality is stark: no reliable internet access, intermittent grid power, and zero access to datacenter accelerators.
In our paper presented at the 7th IEEE International Conference on Recent Advances in Intelligent Computational Systems (RAICS 2026), Prof. Rajeev Rajan and I investigated a practical engineering question: can an edge-optimized neural network with under 4 million parameters match heavyweight baselines for rare dermatological diagnosis?
The Clinical Context: Harlequin Ichthyosis
Harlequin Ichthyosis is an extraordinarily severe congenital skin disorder caused by mutations in the ABCA12 gene. Early clinical recognition is essential for neonatal survival, yet primary physicians in rural clinics often lack specialized dermatological training. Automated diagnostic assistance must operate directly on low-power handhelds and embedded diagnostic probes without round-trip cloud latency.
Architectural Design: HI-MobileNet
We evaluated multiple convolutional backbones and identified MobileNetV2 with inverted residual bottlenecks and depthwise separable convolutions as the optimal Pareto boundary between compute and feature discrimination:
- Depthwise Separable Convolutions: Factorizes standard spatial convolution into lightweight depthwise spatial filtering followed by $1 \times 1$ pointwise projection, reducing FLOPs by 8.5×.
- Inverted Residual Bottlenecks: Maintains thin bottleneck interfaces while expanding to higher-dimensional manifolds internally, preventing non-linear ReLU layers from destroying subtle epidermal texture information.
- Targeted Transfer Learning: Pre-trained ImageNet feature extractors fine-tuned with progressive unfreezing and discriminative learning rates.
Empirical Results
| Model | Parameters | Accuracy | Inference (CPU Edge) |
|---|---|---|---|
| ResNet-50 | 25.6 M | 98.84% | 112 ms |
| DenseNet-121 | 8.0 M | 99.12% | 74 ms |
| BiomedCLIP | 21.1 M | 96.50% | 185 ms |
| HI-MobileNet (Ours) | 3.4 M | 99.96% | 14.2 ms |
Key Takeaways for Practitioners
- Domain-specific data curation beats parameter count: Rigorous preprocessing and artifact cleanup yielded far larger performance gains than increasing model depth.
- Inverted residuals preserve granular medical textures: High-frequency diagnostic cues (keratin fissures, scale boundaries) are easily blurred by aggressive spatial pooling.
- Edge feasibility is achievable: Modern low-power architectures can deliver institutional-grade diagnostic assistance at a fraction of the hardware cost.