Ensemble Multiple Instance Learning for Bladder Cancer Diagnosis

This paper presents a method for automated diagnosis of bladder cancer from digital cytology slides. The proposed method relies on a-priori selection of the most atypical cells and an ensembling of Multiple Instance Learners to predict diagnosis. Our model is trained on a large clinical trial dataset to predict the outcome of cystoscopy and histology examinations directly from voided urine cytology slides. To the best of our knowledge, it is the first time such approach is published. The considered task is known difficult: Yafi et al. [1] evaluated the sensitivity and specificity of experts analysing voided urine cytology respectively at 30% and 87%. The proposed method achieves a sensitivity of 76% and a specificity of 79%, showing that computer-aided analysis of digital cytology slides is a promising approach for bladder cancer diagnosis able to improve patient care as, unlike cystocopy/histology examinations, voided urine cytology is non-invasive and inexpensive.