UM researchers develop AI to distinguish between renal cancer subtypes
Renal cell carcinoma (RCC) is among the top ten most common cancers worldwide. However, approximately 20% of small renal masses removed in surgery are found to be benign. Therefore, a considerable number of patients undergo unnecessary surgery with all its associated risks. Furthermore, a meta-analysis found that one in four renal biopsies reported as a benign mass was found to be malignant following surgery. These patients risk late diagnosis and its consequential poorer prognosis.
For patients and clinicians, a better pre-surgical diagnostic tool would be invaluable. Dr Sabrina Rossi and her team are proposing just that with MethylBoostER (Methylation and XGBoost for Evaluation of Renal tumors).
Utilising DNA methylation data, researchers trained multiclass XGBoost
machine learning models to classify four types of renal tumour and normal tissue. Over training and testing, the model accurately distinguished between the four pathological tumour types in 96% of samples analysed. Although three benign samples were among those incorrectly predicted, no malignant samples were miscategorised as benign.
To help with clinician decision-making, Dr Rossi and her team built in a confidence threshold for predictions (output probability larger than 0.85), under which the model produces two predictions. For high-confidence predictions, the model's accuracy in testing was 0.982 with the moderate-confidence (<0.85) predictions scoring 0.871 accuracy if the model's first and second predictions are considered.
Due to its ease of use, researchers envision that MethylBoostER could be used in clinical settings in the future after further refinement with larger data sets.
Co-author Izzy Newsham has created an excellent summary of the paper with figures here, this thread includes links to the project's data and code.
You can read the full article here: S. Rossi et al., Accurate detection of benign and malignant renal tumor subtypes with MethylBoostER: An epigenetic marker–driven learning framework. Science Advances 8 (39), 2022.
H. D. Patel, S. C. Druskin, S. P. Rowe, P. M. Pierorazio, M. A. Gorin, M. E. Allaf, Surgical histopathology for suspected oncocytoma on renal mass biopsy: A systematic review and meta-analysis. BJU Int.119, 661–666 (2017).