An automatic AI model for tumoral burden detection and segmentation on whole-body MRI: The DIPCAN study.

e13717 Background: Metastasis dissemination is the main cause of death from cancer, being whole-body magnetic resonance imaging (MRI) a key technology for lesion detection and staging. However, radiological reading is limited by the lack of efficient, accurate tools to automate these tasks, hindering its widespread clinical adoption. As part of the DIPCAN study, focused on generating artificial intelligence (AI) models to prevent, diagnose, and ease the treatment choice of metastatic cancer patients, we aimed to develop an automatic detection and segmentation algorithm to extract their heterogeneity and patient phenotype on whole-body MR exams. Methods: Whole-body MR exams from 518 metastatic pan-cancer patients with head and neck, breast, ovary, uterus, prostate, colorectal, gastric, thymus, lung, esophagus, liver and biliary tract, pancreas, pleura or peritoneum, sarcoma, thyroid, kidney, bladder or urinary tract primary tumors were prospectively collected. Patient scans were acquired using the same protocol on two different scanners (1.5 and 3T) at Hospital MD Anderson Cancer Center (MDACC), Madrid, Spain. Transversal diffusion weighted imaging (DWI) MR sequences from five segments (head-chest-abdomen-pelvis-thighs) were selected for model development. The exams were read and lesion detection was informed by five experienced radiologists from MDACC. Manual segmentations were performed by expert image technicians. A total of 669 segments from 432 patients including 1417 lesions were used for training, while 134 segments from 86 patients including 284 lesions were reserved for testing. The 3D full resolution nn-Unet v2 was implemented following a 5-fold cross-validation (CV) strategy. For the inference process and testing, models resulting from each fold in the CV, as well as multiple ensemble combinations were evaluated. To assess detection performance, sensitivity was calculated, while for segmentation validation, the average lesion Dice coefficient was calculated on correctly detected lesions. Results: The best model from the CV provided the highest detection and segmentation performance in the validation cohort, with a sensitivity (se.) of 0.64 and a lesion average Dice coefficient of 0.78. These results overcome the current state-of-the-art metrics, which reports a se. of 0.63 and lesion average Dice coefficient of 0.53. Conclusions: The AI model implemented allowed the successful detection and segmentation of tumoral lesions in whole-body MR scans, outperforming current methods. This methodology could streamline workflows for radiologists and oncologists, by enhancing quantitative information extraction from regions of interest, while providing added value to clinical diagnosis.