Patient-Specific Deep Learning Model for Clinical Target Volume Delineation on Daily CBCT of Breast Cancer Patients based on Intentional Deep Overfit Learning (IDOL) Framework

Abstract

Increasingly complex target volumes and the use of modern irradiation techniques emphasize the importance of daily image guidance more than ever. Significant progress has been made in adjuvant breast cancer radiotherapy (RT) and the need for optimized image guidance is growing. Furthermore, the position of the breast during RT after breast-conserving surgery is highly variable than expected. In this context, cone beam computed tomography (CBCT) is a very effective tool enabling prompt and accurate adaptive radiation therapy (ART). In this study, we aim to develop a deep learning (DL)-based algorithm to segment clinical target volume (CTV) from daily CBCT scans. Also, we validate the optimization of further learning when applying the Intentional Deep Overfit Learning (IDOL) framework. A total of 240 different CBCT scans obtained from 100 breast cancer patients were used for this study. CTV was defined as whole breast plus margin in all patients. The workflow consists of two training stages: (1) training a novel 'generalized' DL model (Swin_UNETR) to identify and delineate breast CTV on CBCT scans using 90 breast cancer patient cases (2) applying an 'intentional overfitting' to the 'generalized' DL model to generate a 'patient-specific' model using the remaining 10 breast cancer patients. In this study, for the intentionally overfitting stage, we additionally trained with CBCT scans from the patient's 1st fraction to the 14th fractions cases. The results of the proposed method were compared quantitatively with the expert's contours on 1st-15th fractions CBCT scans using Dice Similarity Coefficient (DSC). The average DSC between the 'generalized' DL model-based breast CTV contours and reference contours for the patient's 15th fraction was 0.9672. When implementing the IDOL framework with the CBCT scan obtained during the patient's 1st treatment, the average DSC was improved to 0.9809. When additional CBCT scans taken during each of the 1st to 6th fractions were used for training, the average DSC could be most effectively raised to 0.9835. The p-value comparison between the 'generalized' DL model and the 1st fraction was found to be 3.62E-04, while the comparison with the 6th fractions resulted in a p-value of 8.36E-05. The average time required for IDOL training using one CBCT scan and six CBCT scans was 107 seconds and 127 seconds, respectively. In this study, we developed a patient-specific DL-based training algorithm to segment CTV in CBCT scans for breast cancer patients. The performance improvement was relatively significant and was confirmed that using continual DL with additional CBCT scans, which are taken every day, can be more accurate and efficient than drawing breast CTV using a general model. Our novel patient-specific model can be effectively applied to various ARTs by not only reducing labor and time but also increasing accuracy.