Exploring Feature Selection with Deep Learning for Kidney Tissue Microarray Classification Using Infrared Spectral Imaging.

Kidney and renal pelvic cancer are a significant cause of cancer-related deaths, with the most common malignant kidney tumor being renal cell carcinoma (RCC). Chromophobe renal cell carcinoma is a rarer form of RCC that poses significant challenges to accurate diagnosis, as it shares many histologic features with Oncocytoma, a benign renal tumor. Biopsies for histopathological and immunohistochemical analysis have limitations in distinguishing chromophobe RCC from Oncocytoma. Syndromic cases may also have tumors with overlapping features. Techniques such as infrared (IR) spectroscopic imaging have shown promise as an alternative approach to tissue diagnostics. In this study, we propose a deep-learning-based framework for automating classification in kidney tumor tissue microarrays (TMAs) using an IR dataset. Feature selection algorithms reduce data dimensionality, followed by a deep learning classification approach. A classification accuracy of 91.3% was observed for validation data, even with the use of 13.6% of all wavelengths, thereby reducing training time by 21% compared to using the entire spectrum. Through the integration of scalable deep learning models coupled with feature selection, we have developed a classification pipeline with high predictive power, which could be integrated into a high-throughput real-time IR imaging system. This would create an advanced diagnostic tool for the detection and classification of renal tumors, namely chromophobe RCC and Oncocytoma. This may impact patient outcomes and treatment strategies.