Addressing biases in gastric cancer diagnosis through generative models and vision-based surface tactile sensing
Siddhartha Kapuria, Naruhiko Ikoma, Sandeep Chinchali, Farshid Alambeigi · 2025
Towards developing an informed, intuitive, and generalized artificial intelligence model for endoscopic diagnosis of Advanced Gastric Cancer (AGC) lesions, in this work, we extend the use of generative models to augment limited and biased training datasets. This approach aims to enhance model performance and generalizability, tackling the challenge of acquiring sizable and well-balanced datasets in the clinical domain. We applied this approach to our unique dataset comprised of textural images of additively-manufactured realistic AGC tumor phantoms collected using a robotic data collection system and our unique Vision-based Tactile Sensor - HySenSe. Our approach was particularly suited to this dataset, since HySenSe’s limited sensing area captures only partial images of the much larger AGC lesions, which reduces the amount of information available in a single image. We trained a single Class-Conditioned Latent Diffusion Model (CC-LDM) on an artificially imbalanced dataset of HySenSe images, logging performance using Fr´echet Inception Distance (FID), and exploring different hyperparameter configurations. Using the high-quality synthetic images generated using this model, we performed experiments using our previously developed Dilated ResNet model for AGC tumor classification with the following objectives: (i) study the amount of synthetic data required during training, and (ii) compare different data addition strategies during model training using cross-validation. Note that while the models were trained using both real and synthetic textural images, testing was performed solely on real images. Our experiments showed that the trained classifier demonstrates improved generalizability when trained on a mixture of real and synthetic images, but not all methods of mixing the data are equally effective. Overall, the model shows improved performance even under mixed morphological tumor conditions and partial sensor contact.