Radiological staging clinical decision support model for rectal cancer lymph node metastasis detection on MRI

BACKGROUND: Accurate staging of lymph node metastasis (LNM) is crucial for personalising rectal cancer treatment. Lymph nodes (LNs) are the most common sites of rectal cancer metastasis, and malignant LNs are typically treated with neo-adjuvant radiotherapy or chemoradiotherapy (CRT) to reduce the chance of recurrence and distant metastasis after surgery. Radiological staging criteria, based on LN size, shape, and texture, are known to be subjective, and radiologists have an average diagnostic performance of 73% sensitivity and 74% specificity. METHODS: This study develops a fully automatic and end-to-end pre-operative radiological staging model for rectal cancer LNM using artificial intelligence (AI) methods. The model combines automatic detection of lymph nodes on Magnetic Resonance Imaging (MRI) with multiple instance learning patient-level lymph node staging. Models were trained and evaluated on an in-house dataset provided by Leeds Teaching Hospitals NHS Trust (LTHT) including 458 patients with pre-operative MRI scans, patient clinical data, and post-operative pathological TNM staging. Accurate detection of the lymph nodes on MRI is achieved using nnU-Net, and the classification results compare different 3D feature encoders, investigating a trade-off between performance and interpretability. RESULTS: The results demonstrate state-of-the-art performance with cross-validated metrics of 0.828 AUC, 86.6% sensitivity, and 72.4% specificity. Clinical validation study results show that the AI staging model performance exceeded three expert radiologists in predicting post-operative pathology with a 9% higher F1 score. CONCLUSIONS: This study provides strong evidence that an end-to-end AI model can significantly improve pre-operative staging of rectal cancer lymph node metastasis. The model outperformed expert radiologists and shows clear potential to enhance clinical decision making and improve patient outcomes.