KS-TMIL: A K-Stage Transformer approach with multiple instance learning model for ovarian cancer subtype classification
Existing multiple instance learning (MIL) methods treat whole slide image (WSI) as collections of independent patches, neglecting these crucial spatial and morphological connections. This paper proposes K-Stage Transformer with multiple instance learning framework (KS-TMIL) that leverages the power of transformers. Unlike traditional MIL methods, KS-TMIL excels at modeling relationships between patches within a WSI. This capability is achieved through a multi-stage transformer that incorporates inception based position encoding generator (IBPEG) block. This enables KS-TMIL to effectively leverage the weakly supervised labels and achieve accurate classifications. KS-TMIL offers a family of models with varying numbers of transformer layers, catering to different classification tasks. The evaluation is conducted using publicly available Kaggle dataset "UBC ovarian cancer subtype classification and outlier detection", demonstrating the effectiveness of the proposed model with an AUC and balanced accuracy of 99% and 93.3% respectively.