Development of Artificial Intelligence (AI) Based Platform for Locally Advanced Rectal Cancer Prognosis
Y Zhang1*, L Shi2, X Sun2, S Jabbour1, N Yue1, K Nie1, (1) Rutgers Cancer Institute of New Jersey, New Brunswick, NJ, (2) Department of Radiation Oncology, Sir Run Run Shaw Hospital, Zhejiang Univ., Hangzhou, China
Presentations
TH-D-TRACK 3-4 (Thursday, 7/29/2021) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Purpose: To develop a deep learning based platform to explicitly quantify multi-regional phenotypes of local advanced rectal cancer (LARC) and assess the association with neoadjuvant chemoradiation (CRT) outcome.
Methods: A total of 203 patients receiving neoadjuvant CRT followed by total mesorectal excision (TME) were enrolled. All underwent anatomical T1/T2, diffusion-weighted MRI (DWI) and dynamic contrast-enhanced (DCE) MRI before CRT. 133 cases were used for training with four-fold validation, while remaining 70 cases treated at a later time was used as a separate validation set. Pixel-based deep learning algorithm with convolutional neural network (CNN) was developed to automatically localize the rectum and further identify tumor based on all-frame DCE sequences. A radiomics sequencer was then constructed using deep-learning based approach to understand tumor phenotypes across multi-modality MRI images. The radiomics sequencer was further utilized to associate with treatment outcome as pathologic complete response (pCR), good response (GR), and T-downstaging. Results were evaluated using area under the receiver operating characteristics curves (AUC).
Results: The segmentation result was evaluated by comparing with physicians’ contours using Dice Similarity Coefficient (DSC) with a mean of DSC is 0.84 among all cases. A deep-learning based radiomics sequencer was built in predicting pCR, GR and T-downstaging, with AUCs of 0.82, 0.79, and 0.77 respectively. It is also found that DCE sequences especially the second phase of post-contrast T1 contributed the most predictive features followed by DWI sequence.
Conclusion: Based on the radiomics features from multiparametric MR imaging, machine learning is a feasible way to localize and predict the pathological outcome. Especially, support vector machines can provide better performance than other simple models.
Handouts
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email