Toward unbounded open-set recognition to say "I don't know" for glomerular multi-lesion classification

Abstract

Glomeruli are histological structures located at the beginning of the nephrons in the kidney, having primary importance in diagnosing many renal diseases. Classifying glomerular lesions is time-consuming and requires experienced pathologists. Hence automatic classification methods can support pathologists in the diagnosis and decision-making scenarios. Recently most of state-of-the-art medical imaging classification methods have been based on deep-learning approaches, which are prone to return overconfident scores, even for out-of-distribution (OOD) inputs. Determining whether inputs are OOD samples is of underlying importance so as to ensure the safety and robustness of critical machine learning applications. Bare this in mind, we propose a unified framework comprised of unbounded open-set recognition and multi-lesion glomerular classification (membranous nephropathy, glomerular hypercellularity, and glomerular sclerosis). Our proposed framework classifies the input into in- or OOD data: If the sample is an OOD image, the input is disregarded, indicating that the model “doesn’t know” the class; otherwise, if the sample is classified as in-distribution, an uncertainty method based on Monte-Carlo dropout is used for multi-lesion classification. We explored an energy-based approach that allows open-set recognition without fine-tuning the in-distribution weights to specific OOD data. Ultimately, our results suggest that uncertainty estimation methods (Monte-Carlo dropout, test-time data augmentation, and ensemble) combined with energy scores slightly improved our open-set recognition for in-out classification. Our results also showed that this improvement was achieved without decreasing the 4-lesion classification performance, with an F1-score of 0.923. Toward an unbounded open-set glomerular multi-lesion recognition, the proposed method also kept a competitive performance.

Ähnliche Arbeiten