Uncertainty-aware Renal Cell Carcinoma Subtype Classification

Kidney cancer requires accurate and timely diagnosis to guide effective treatment as a particularly prevalent cancer in older adults. Early detection and the precise identification of cancer subtypes and stages can significantly influence therapeutic decisions which leads to improved patient outcomes, preventing metastasis, and increasing survival rates. Renal cell carcinoma (RCC), the most common form of kidney cancer, is a heterogeneous type of cancer that comprises several subtypes, including clear cell RCC, papillary RCC, chromophobe RCC, and oncocytoma. Each of these subtypes comes with a unique biological behavior and treatment response. Currently, clinicians adopt a step-by-step approach to classify these subtypes based on the level of diagnostic uncertainty, starting with tumor morphology. While tumor morphology serves as the initial method for evaluation, its reliance on overlapping features, such as similar cellular structures and staining patterns, often introduces ambiguity and complicates accurate classification. This method is also a time-intensive process which requires significant expertise. When uncertainty persists during tumor morphology analysis, clinicians proceed to more advanced techniques like immunohistochemical (IHC) profile analysis, which, although valuable, comes with its own set of challenges, including high costs and the need for specialized expertise. Together, these factors add complexity to clinical decision-making and extend the diagnostic timeline. Recent advancements in deep learning, particularly convolutional neural networks (CNNs), have opened new avenues for enhancing cancer diagnosis. CNNs are adept at processing histopathological images due to their ability to capture complex spatial patterns. This capability enables the differentiation of various cell structures and tissue textures, which are critical elements in the accurate classification of cancer subtypes. CNNs are less computationally intensive and work more effectively with smaller datasets, which are common in medical imaging. This makes them a better fit for tasks in this field than more complex models like Transformers. In our work, we leverage the strengths of supervised models, which benefit from well-labeled data, allowing for precise pattern recognition and reducing the margin for error. While self-supervised and weakly-supervised methods can extract valuable patterns from vast amounts of data with minimal labeling, their performance may fall short when distinguishing between subtle differences required for accurate medical diagnoses that often demand highly detailed and specific labeling. Building on these advancements, we propose a hybrid model that incorporates deep learning for the initial detection of tumor regions and subtype classification. In cases where the model exhibits uncertainty in the initial classification, it triggers a secondary validation step using traditional machine learning techniques applied to immunohistochemistry (IHC) analysis for more accurate confirmation. This integrated approach allows for a more comprehensive diagnostic framework, merging morphological insights from histopathology with molecular data from IHC. As a result, our model enhances RCC subtype classification accuracy while reducing processing time and cost, offering a promising solution for improving diagnostic precision in clinical practice.