SGProtoNet: Semantic-Guided Prototypical Networks for Multi-Label Few-Shot Medical Image Classification
ISEF · 2026 Robotics and Intelligent Machines
Overview
Automated analysis of medical images offers transformative potential for clinical diagnostics, yet two persistent obstacles limit practical adoption: the high cost of expert annotation and the inherent multi-label nature of radiological findings. Acquiring sufficient labeled data for deep learning remains especially difficult when pathological categories are rare or previously unencountered. To tackle these issues, we introduce SGProtoNet, a Semantic-Guided Prototypical Network that unifies few-shot learning with multi-label chest X-ray classification using the IU-CXR dataset. Rather than relying solely on visual supervision, SGProtoNet grounds its prototype representations in paired radiological reports, exploiting the descriptive power of clinical language to compensate for visual data scarcity. Concretely, frozen ViT-B/16 and PubMedBERT encoders first bring image patches and report tokens into a common embedding space through contrastive pre-training; a Semantic-Guided Attention Module (SGAM) then cross-attends over patch features with report embeddings as queries, yielding semantically enriched visual representations that are fused via a learnable gating mechanism. The resulting prototypes are subsequently employed within a 5-way K-shot episodic meta-learning scheme, enabling the network to recognize unseen findings from as few as one to five annotated examples. Comprehensive experiments yield a Macro AUC of 90.6% and a Macro F1-Score of 0.75, surpassing prior baselines and establishing semantically guided prototypical learning as a compelling strategy for data-scarce, multi-label medical image recognition.
Awards (2)
- King Abdulaziz & his Companions Foundation for Giftedness and Creativity: Full Scholarship from King Fahd University of Petroleum and Minerals(KFUPM) (and a $400 cash prize) $400
- King Abdulaziz & his Companions Foundation for Giftedness and Creativity: NOT TO BE READ -- $400 cash prize for each Full Scholarship from King Fahd University award recipient $400
Competition history
- ISEF 2026
Resources
Related projects
ISEF · 2025
Evaluating Convolutional Neural Networks for Multi-Label Chest X-ray Diagnosis: Model Complexity, Hardware Efficiency, and Radiologist Comparison
ISEF · 2021
Using A Novel Semi-Supervised Machine Learning Method to Improve Image-based Lung Cancer Diagnostic Algorithms
ISEF · 2022
Screen CXR: A Novel Deep Learning-Based Multi-Model Pipeline for Detection of Any Lung Tissue Disease Through Automatic Chest X-Ray Image Analysis
ISEF · 2025
M³LungXAI-LF: An Explainable Multi-Modal Late Fusion System for Improved Clinical Early Lung Cancer Diagnosis With Automated Medical Reporting
Closest projects by meaning, across every fair and year in the corpus.
Source: Regeneron International Science and Engineering Fair