Quantifying Biases in Vision-Based Surface Tactile Sensing Gastrointestinal Cancer Datasets using Data Valuation

Current approaches for diagnosing cancers of the gastrointestinal (GI) tract using artificial intelligence are limited by the scarcity and inherent biases of medical data, which lead to missed diagnoses. While conventional data augmentation can address class imbalance, it often fails to improve underlying data quality or diversity. To address these limitations, this work introduces a data-valuation framework to understand the intrinsic value of each data point. Leveraging a novel vision-based tactile sensor (VTS) capable of capturing high-resolution textural images, we employ data valuation to quantify the contribution of each sample to the downstream GI tumor classification task. We benchmark two complementary methods, Data Valuation via Gradient Similarity (DVGS) and TracIn, to assign a “worth” to each textural image and validate these scores through targeted data removal experiments. Our results demonstrate that data valuation effectively stratifies our dataset. The methods consistently assign high value to rare and morphologically distinct samples while identifying redundancy in more common classes. We demonstrate that removing these high-value samples leads to a significantly sharper degradation in both predictive accuracy and inter-class fairness compared to random removal. Using the same valuation approach, we examine the inductive biases of diverse architectures and uncover how convolutional and transformer-based networks prioritize different features within the same dataset. Our results show that agreements between valuation methods and models vary substantially, revealing architecture-dependent sensitivities to texture and surface contact. This study establishes a principled methodology for dataset curation based on quantified sample values. By identifying the characteristics of the most valuable data, we also provide a clear pathway towards guiding generative models in synthesizing data that is demonstrably effective, representing a crucial step toward engineering more robust, data-efficient, and equitable AI systems for critical medical applications.