TLDR: FOSSIL is a new AI framework that uses regret-minimizing curriculum learning to improve Mpox diagnosis, especially with limited and imbalanced data. It adaptively weights training samples based on their difficulty, determined by the model’s own uncertainty. This approach significantly enhances diagnostic accuracy, calibration, and robustness across various deep learning architectures, making AI models more reliable and interpretable for real-world medical applications like teledermatology, without needing extensive metadata or synthetic data augmentation.
In the rapidly evolving field of medical diagnostics, deep learning holds immense promise, especially for identifying diseases like Monkeypox (Mpox). However, a significant hurdle remains: training robust AI models with limited, imbalanced, and often metadata-free datasets. Traditional methods often struggle with unstable optimization and poor generalization when data is scarce, leading to unreliable diagnoses.
Addressing this critical challenge, a new framework called FOSSIL (Flexible Optimization via Sample-Sensitive Importance Learning) has been introduced. This innovative approach offers a regret-minimizing weighting system that intelligently adjusts how much emphasis is placed on different training samples based on their perceived difficulty. This adaptive learning process is particularly beneficial for biomedical datasets where obtaining large, diverse samples can be difficult due to privacy concerns, cost, or the rarity of the disease itself.
The FOSSIL framework integrates with both convolutional and transformer-based AI architectures, creating a four-stage curriculum (Easy–Very Hard) for learning. It uses a clever method to determine sample difficulty: by analyzing the model’s own uncertainty in its predictions, specifically using softmax-based confidence. This means the model itself helps decide which examples are “easy” or “hard” to learn from, rather than relying on manual annotations or expert heuristics which can be subjective and limit the model’s adaptability.
The researchers applied FOSSIL to the diagnosis of Mpox skin lesions, a disease that presents diagnostic challenges due to its visual similarity to other conditions. In telemedicine or low-resource settings, crucial patient information or advanced lab tests are often unavailable, making accurate, metadata-free AI models highly valuable. The study utilized publicly available datasets like the Mpox Skin Lesion Dataset Version 2.0 (MSLD v2) for training and internal validation, and the Mpox Close Skin Images (MCSI) dataset for external validation, ensuring a rigorous evaluation of the framework’s real-world applicability.
A key finding was FOSSIL’s ability to significantly improve diagnostic performance. Across various model architectures, FOSSIL consistently enhanced discrimination, calibration, and robustness, even when faced with real-world image distortions like blur or compression. For instance, the ConvNeXt-T model, when trained with FOSSIL, achieved an impressive AUC of 0.9573, demonstrating superior performance without showing signs of overfitting. This means the model not only performs well on the data it was trained on but also generalizes effectively to new, unseen cases.
Beyond just improving accuracy, FOSSIL also contributes to the interpretability and trustworthiness of AI models. By understanding which parts of an image the model focuses on (through techniques like Score-CAM), researchers observed that in “easy” cases, the model’s attention was sharply localized around the lesion. In “very hard” cases, where visual ambiguity was high, the attention was more diffuse but still structured, reflecting the model’s intrinsic uncertainty in differentiating visually similar patterns. This alignment of machine perception with clinical reasoning is crucial for deploying AI in sensitive healthcare environments.
Also Read:
- Learning from Mistakes: How Language Feedback Boosts AI in Embodied Tasks
- ViTAR: A New AI Model for Iterative Visual Reasoning in Medical Imaging
The implications of FOSSIL extend far beyond Mpox diagnosis. The framework’s unified mathematical structure allows it to be applied to a wide range of biomedical domains, including radiology, histopathology, microscopy, and genomics, where data scarcity and class imbalance are common. By reducing reliance on extensive data augmentation and metadata, FOSSIL offers a generalizable, data-efficient, and interpretable solution for developing reliable AI in medical imaging under data-limited conditions. This research paves the way for a new generation of AI systems that are not only accurate but also transparent and trustworthy in supporting clinical decisions. You can read the full research paper here: FOSSIL: Regret-minimizing curriculum learning for metadata-free and low-data Mpox diagnosis.
