| Popis: |
Purpose: To develop and validate a convolutional neural network (CNN) to identify regions of interest (ROIs) suspicious for ductal carcinoma in situ (DCIS) and residual malignancy in lumpectomy margins using wide-field optical coherence tomography (WF-OCT). Background: WF-OCT is the optical analog of high-frequency ultrasound and produces high-resolution intraoperative imaging in real time, with a tissue penetration depth up to 2 mm. Multi-reader studies of WF-OCT have demonstrated the ability to differentiate normal breast parenchyma from neoplasms with greater than 85% sensitivity and specificity. Intraoperative evaluation of lumpectomy specimens using WF-OCT may aid in achieving negative margins at the time of primary surgery and avoid re-excisions. CNNs, a form of artificial intelligence (AI), can be trained to spot ROIs in WF-OCT images of margins suspicious for DCIS and, more generally, residual malignancy. Methods: Lumpectomy margins from 126 patients with ductal malignancy were imaged using WF-OCT, compared to permanent histology (PH), and annotated by board-certified breast pathologists to create a training set of 25,000 control ROIs. A CNN algorithm was developed with 3 convolutional layers, a 3x3 kernel, and 3 fully connected layers to perform binary classification of images as either “suspicious” or “non-suspicious” for malignancy. A weighted loss function was implemented to balance the training data available for non-suspicious vs. suspicious images and to tune sensitivity and specificity. Once trained and properly weighted, the CNN was tested in a prospective study using WF-OCT images of margins from 29 lumpectomy specimens from 29 patients with biopsy-proven DCIS, invasive ductal carcinoma (IDC), or both. The CNN results were compared to PH. Results: Patients were 61.5 ± 7.3 years old, 100% female, with Stage 0-1 disease. Disease types included DCIS (n=27), atypical ductal hyperplasia (n=24), IDC (n=20), invasive lobular carcinoma (n=2), mixed (n=74), and benign findings including usual ductal hyperplasia (n=35), atypical lobular hyperplasia (n=19), duct ectasia (n=17), lymphatic invasion (n=13), and lobular carcinoma in situ (n=12). Following primary surgery, fresh margins were scanned using WF-OCT and approximately 1.9M ROIs were analyzed by the CNN, yielding 15,136 as suspicious for malignancy. Overall, four hundred and ten (410) ROIs were correctly identified, yielding a 74% true positive and 0.8% false positive detection rate; sensitivity and specificity were 74.4% and 99.2%, respectively. Specific to DCIS, the CNN demonstrated a 73% true and 0.5% false positive rate; sensitivity and specificity were 73.0% and 99.5%, respectively. Conclusions: Automated analysis of WF-OCT images of lumpectomy specimens, using a trained CNN to identify ROIs suspicious for malignancy is feasible, demonstrating high concordance with PH. Specific to DCIS, the CNN demonstrated equivalent utility with a lower false positive rate. A prospective trial is needed to evaluate specimens in real time to determine improvement in re-excision rates. Citation Format: David Rempel, Andrew Berkeley, Alastair Thompson, Savitri Krishnamurthy, Beryl Augustine, Kelly Hunt, Ismail Jatoi, Alia Nazarullah, Chandandeep Nagi, Yanir Levy. Development and validation of a convolutional neural network to identify ductal carcinoma in situ in lumpectomy margins using wide field optical coherence tomography [abstract]. In: Proceedings of the AACR Special Conference on Rethinking DCIS: An Opportunity for Prevention?; 2022 Sep 8-11; Philadelphia, PA. Philadelphia (PA): AACR; Can Prev Res 2022;15(12 Suppl_1): Abstract nr A004. |
