Ultrasound-based deep learning radiomics in the assessment of pathological complete response to neoadjuvant chemotherapy in locally advanced breast cancer

• Published in: European journal of cancer (1990) Vol. 147; pp. 95 - 105
• Main Authors: Jiang, Meng, Li, Chang-Li, Luo, Xiao-Mao, Chuan, Zhi-Rui, Lv, Wen-Zhi, Li, Xu, Cui, Xin-Wu, Dietrich, Christoph F.
• Format: Journal Article
• Language: English
• Published: England Elsevier Ltd 01-04-2021; Elsevier Science Ltd
• Subjects: Algorithms; Biopsy; Breast cancer; Calibration; Chemotherapy; Confidence intervals; DDeep learning; Decision analysis; Deep learning; Epidermal growth factor; ErbB-2 protein; Growth factors; Locally advanced breast cancer; Machine learning; Nomograms; Phenotypes; PPathological complete response; Radiomics; Receptors; Redundancy; Risk analysis; Risk factors; RRadiomic nomogram; Statistical analysis; Subgroups; Tumors; Ultrasonic imaging; Ultrasonic testing; Ultrasound; RRadiomic nomogram; Locally advanced breast cancer; PPathological complete response; DDeep learning
• ISSN: 0959-8049; 1879-0852
• DOI: 10.1016/j.ejca.2021.01.028

The aim of the study was to develop and validate a deep learning radiomic nomogram (DLRN) for preoperatively assessing breast cancer pathological complete response (pCR) after neoadjuvant chemotherapy (NAC) based on the pre- and post-treatment ultrasound.
Patients with locally advanced breast cancer (LABC) proved by biopsy who proceeded to undergo preoperative NAC were enrolled from hospital #1 (training cohort, 356 cases) and hospital #2 (independent external validation cohort, 236 cases). Deep learning and handcrafted radiomic features reflecting the phenotypes of the pre-treatment (radiomic signature [RS] 1) and post-treatment tumour (RS2) were extracted. The minimum redundancy maximum relevance algorithm and the least absolute shrinkage and selection operator regression were used for feature selection and RS construction. A DLRN was then developed based on the RSs and independent clinicopathological risk factors. The performance of the model was assessed with regard to calibration, discrimination and clinical usefulness.
The DLRN predicted the pCR status with accuracy, yielded an area under the receiver operator characteristic curve of 0.94 (95% confidence interval, 0.91–0.97) in the validation cohort, with good calibration. The DLRN outperformed the clinical model and single RS within both cohorts (P < 0.05, as per the DeLong test) and performed better than two experts' prediction of pCR (both P < 0.01 for comparison of total accuracy). Besides, prediction within the hormone receptor–positive/human epidermal growth factor receptor 2 (HER2)–negative, HER2+ and triple-negative subgroups also achieved good discrimination performance, with an AUC of 0.90, 0.95 and 0.93, respectively, in the external validation cohort. Decision curve analysis confirmed that the model was clinically useful.
A deep learning–based radiomic nomogram had good predictive value for pCR in LABC, which could provide valuable information for individual treatment.
•A novel preoperative pathological complete response prediction model was developed.•The model is based on pre- and post-neoadjuvant chemotherapy ultrasound images.•It yielded an area under the receiver operator characteristic curve AUC of 0.94 in the independent external validation cohort.•It outperformed two experts who evaluated the pathological complete response status.•It may facilitate tailoring the optimum extent of breast and axillary surgery.