A Hybrid Deep Learning Guided Few-Shot Biomedical Image Classification for Resource Constrained Clinical Settings

The insufficient annotated data, great variability of modalities, and the computational requirements of standard deep learning models are a strong impediment to biomedical image classification in low-resource clinical environments. This paper proposes a hybrid few-shot learning model, a CNN-Transformer-based feature extractor with Prototypical Networks and embedding refinement techniques addressing them. The model is trained on fewshot tasks designed based on the ChestX-ray14 dataset, where the conditions are realistic low-data episodes of 5way, 5 -shot. It combines feature hallucination to artificially boost sparse classes and employs adversarial domain adaptation to boost cross-modality (X-ray and MRI) generalizability. The accuracy of the proposed model reaches 95.08, which is much higher than a number of state-of-the-art baselines, such as Matching Networks, MAML, and standalone Transformer or CNN models. Assessment involves the performance in various shots, domains and model ablations, which exhibit strong generalization and performance. The framework is lightweight and can run on resource-constrained edge hardware such as Jetson Nano and is able to provide real-time diagnostic assistance to under-serviced areas. This study reveals the opportunities of few-shot hybrid architectures in closing the divide between state-of-the-art AI and access to healthcare to be available to the entire world, thus making it possible to deploy it in realworld clinical environments with a minimum of annotation.