Impact of Transfer Learning Using Local Data on Performance of a Deep Learning Model for Screening Mammography.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To investigate the issues of generalizability and replication of deep learning (DL) models by assessing performance of a screening mammography DL system developed at New York University (NYU) on a local Australian dataset. Materials and Methods In this retrospective study, all individuals with biopsy and surgical pathology-proven lesions and age-matched controls were identified from a South Australian public mammography screening program (January 2010 to December 2016). The primary outcome was DL system performance, measured with the area under the receiver operating characteristic curve (AUC), in classifying invasive breast cancer or ductal carcinoma in situ (n = 425) from no malignancy (n = 490) or benign lesions (n = 44) in age-matched controls. The NYU system, including models without (NYU1) and with (NYU2) heatmaps, was tested in its original form, after training from scratch (without transfer learning; TL), after retraining with TL. Results The local test set comprised 959 individuals (mean age, 62.5 years [SD, 8.5]; all female). The original AUCs for the NYU1 and NYU2 models were 0.83 (95%CI = 0.82-0.84) and 0.89 (95%CI = 0.88-0.89), respectively. When applied in their original form to the local test set, the AUCs were 0.76 (95%CI = 0.73-0.79) and 0.84 (95%CI = 0.82-0.87), respectively. After local training without TL, the AUCs were 0.66 (95%CI = 0.62-0.69) and 0.86 (95%CI = 0.84-0.88). After retraining with TL, the AUCs were 0.82 (95%CI = 0.80-0.85) and 0.86 (95%CI = 0.84-0.88). Conclusion A deep learning system developed using a U.S. dataset showed reduced performance when applied 'out of the box' to an Australian dataset. Local retraining with transfer learning using available model weights improved model performance. ©RSNA, 2024.

Interval cancers as related to frequency of recall to assessment in the South Australian population-based breast screening program: An exploratory study.

Australian accreditation standards specify upper limits for percentages of women recalled for further assessment following screening mammography. These limits have been unchanged since national screening commenced circa 1990, although screening target ages have changed, and technology from analogue to digital mammography. This study compared 2804 women with interval cancers diagnosed since national screening began (cases) with 14,020 cancer-free controls (5 controls per case), randomly selected after matching by age, round, screen type and calendar year of screening episode, to determine the odds of interval cancer by differences in clinic recall to assessment percentages. Within low numbers of recalls that were within accepted accreditation ranges, results did not indicate more frequent recalls to assessment to be associated with fewer interval cancers in the analogue era. However, more frequent recalls were associated with reduced interval cancers for digital screens. These results are not conclusive, requiring confirmation in other screening environments, especially those with larger numbers of digital screens. If confirmed, frequency of recalls to assessment may need adjustment to get the best trade-offs in the digital era between reduced odds of interval cancers from more recalls and increases in financial and non-financial costs, including increased potential for overdiagnosis.

Investigating the breast cancer screening-treatment-mortality pathway of women diagnosed with invasive breast cancer: Results from linked health data.

OBJECTIVE: To examine the screening-treatment-mortality pathway among women with invasive breast cancer in 2006-2014 using linked data. METHODS: BreastScreen histories of South Australian women diagnosed with breast cancer (n = 8453) were investigated. Treatments recorded within 12 months from diagnosis were obtained from linked registry and administrative data. Associations of screening history with treatment were investigated using logistic regression and with cancer mortality outcomes using competing risk analyses, adjusting for socio-demographic, cancer and comorbidity characteristics. RESULTS AND CONCLUSION: For screening ages of 50-69 years, 70% had participated in BreastScreen SA ≤ 5 years and 53% ≤ 2 years of diagnosis. Five-year disease-specific survival post-diagnosis was 90%. Compared with those not screened ≤5 years, women screened ≤2 years had higher odds, adjusted for socio-demographic, cancer and comorbidity characteristics, and diagnostic period, of breast-conserving surgery (aOR 2.5, 95% CI 1.9-3.2) and radiotherapy (aOR 1.2, 95% CI 1.1-1.3). These women had a lower unadjusted risk of post-diagnostic cancer mortality (SHR 0.33, 95% CI 0.27-0.41), partly mediated by stage (aSHR 0.65, 95% CI 0.51-0.81), and less breast surgery (aSHR 0.78, 95% CI 0.62-0.99). Screening ≤2 years and conserving surgery appeared to have a greater than additive association with lower post-diagnostic mortality (interaction term SHR 0.42, 95% CI 0.23-0.78). The screening-treatment-mortality pathway was investigated using linked data.

Soft, Brown Rupture: Clinical Signs and Symptoms Associated with Ruptured PIP Breast Implants.

BACKGROUND: Preoperative signs and symptoms of patients with Poly Implant Prothese (PIP) implants could be predictive of device failure. Based on clinical observation and intraoperative findings 4 hypotheses were raised: (1) Preoperative clinical signs including acquired asymmetry, breast enlargement, fullness of the lower pole, decreased mound projection, and change in breast consistency could be indicative of implant rupture. (2) Device failure correlates with a low preoperative Baker grade of capsule. (3) Brown-stained implants are more prone to implant failure. (4) The brown gel could be indicative of iodine ingression through a substandard elastomer shell. METHODS: Preoperative clinical signs were compared with intraoperative findings for 27 patients undergoing PIP implant explantation. RESULTS: Acquired asymmetry (P = 0.0003), breast enlargement (P = 0.0002), fuller lower pole (P < 0.0001), and loss of lateral projection (P < 0.0001) were all significantly predictive of device failure. Capsule Baker grade was lower preoperatively for ruptured implants. The lack of palpable and visible preoperative capsular contracture could be secondary to the elastic nature of the capsular tissue found. Brown implants failed significantly more often than white implants. Analysis of brown gel revealed the presence of iodine, suggesting povidone iodine ingression at implantation. CONCLUSIONS: Preoperative signs can be predictive of PIP implant failure. Brown-stained implants are more prone to rupture. The presence of iodine in the gel suggests unacceptable permeability of the shell early in the implant's life span. A noninvasive screening test to detect brown implants in situ could help identify implants at risk of failure in those who elect to keep their implants.