Search

Objective This study aimed to clinically validate the performance of a multi-feature AI-based solution on detection of invasive versus in situ carcinomas and in situ ductal carcinoma high grade from atypical ductal hyperplasia/low grade DCIS compared to rigorous ground truth (GT) established by multiple blinded expert pathologists in breast biopsies. Design Performance of the AI solution was prospectively tested on breast biopsies from two medical institutions in different geographies. AI results were compared against the ground truth (GT) established by consensus of two subspecialist breast pathologists. The study endpoints were detection of invasive carcinoma (IDC, ILC) and DCIS/ADH, including differentiating between different DCIS grades and ADH. ADH and DCIS Low Grade were pooled together because of similar clinical management when diagnosed on a biopsy. Results Six pathologists participated in the study and reported on 436 breast biopsies (841 H&E slides), including 156 invasive (including 31 rare subtypes), 135 DCIS/ADH and 145 benign diagnoses. The AI solution demonstrated high performance when compared with the GT with an AUC of 0.99 for the detection of invasive carcinoma (specificity and sensitivity of 93.6% and 95.5% respectively) and with AUC of 0.95 for the detection of DCIS/ADH. The AI solution differentiated well between subtypes/grades of invasive and in-situ cancers with an AUC of 0.97 for IDC vs. ILC and AUC of 0.92 for DCIS high grade vs. low grade/ADH, respectively. Only 11 (7%) cases had discrepancies on invasive diagnosis, 4 of these between invasive versus benign diagnosis encompassing one case on which the invasive component was only represented by rare lympho-vascular invasion, two cases of ILC (one with a diffuse pattern and the second in a case with granulomatous mastitis with multinucleated giant cells and hemosiderin-laden macrophages) and one rare case of tubular carcinoma surrounded by flat epithelial atypia and columnar cell lesions. Fifteen cases (11%) had discrepancies on DCIS/ADH diagnosis between the two specialist pathologists necessitating a third assessment by a specialist to establish GT. Six of these cases also necessitated a review on a multihead microscope to reach a consensus decision, since there was no majority even after 3 reviews. Conclusion This blinded multi-site study reports the successful clinical validation of a multi-feature AI-based solution in detecting and automatically imparting clinically relevant diagnostic parameters regarding invasive and in situ breast carcinoma, offering an important tool for computer-aided diagnosis in routine pathology practice. Citation Format: Anne Vincent-Salomon, Guillaume Bataillon, Alona Nudelman, Judith Sandbank, Anat Albrecht Shach, Lucie Thibault, Lilach Bien, Rachel Mikulinsky, Ira Krasnitsky, Ronen Heled, Chaim Linhart, Manuela Vecsler, Daphna Laifenfeld. A multi-feature AI-based solution for cancer diagnosis in breast biopsies: A prospective blinded multi-site clinical study [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr PD11-04.

Objective This study aimed to clinically validate the use of an AI-based solution by pathologists for the primary diagnosis of breast core needle biopsies as compared with the gold standard practice (review on the microscope). Methods A two-arm prospective reader study comparing the performance of pathologists using an AI-based solution with pathologists using a microscope was performed at two sites (different staining and digital scanners). Both arms were compared to ground truth (GT) established by the consensus of two breast pathologists. Rates of major discrepancies between each arm and GT, as determined by an adjudicating pathologist, were compared. Results Eight pathologists participated in the study and reported on 385 cases (442 HES and 330 H&E slides), each case being reported twice, once in each study arm. Pathologists first reviewed only H&E/HES slides, if requested and available, they were provided with IHCs, while the AI results were on H&E/HES only. The major discrepancy rates of the microscope arm and of the AI arm against GT were 4.42% and 3.12%, respectively, demonstrating a 29.4% reduction in major discrepancies. Pathologists with AI demonstrated very high accuracy for the detection of invasive carcinoma with sensitivity and specificity of 100% for both, as well as for DCIS/ADH with sensitivity of 92.4% and specificity of 97.8%. Conclusions This multi-site reader study reports diagnostic accuracy improvements by pathologists performing diagnosis and reporting with the support of a first read AI solution for breast biopsies. The AI solution performed accurately and generalized well for different staining platforms and different scanners. Thus, AI solutions could be used as significant aiding tools for pathologists in clinical decision-making in routine pathology practice, enhancing the quality and reproducibility of diagnosis. Citation Format: Anne Salomon, Alona Nudelman, Joanna Cyrta, Marina Maklakovski, Anat Albrecht Shach, Geraldine Sebag, Giuseppe Mallel, Ira Krasnitsky, Tali Feinberg, Manuela Vecsler, Judith Sandbank. Primary Diagnosis of Breast Biopsies supported by AI versus Microscope: Multi-Site Clinical Reader Study [abstract]. In: Proceedings of the 2022 San Antonio Breast Cancer Symposium; 2022 Dec 6-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2023;83(5 Suppl):Abstract nr P6-04-07.

There is high demand to develop computer-assisted diagnostic tools to evaluate prostate core needle biopsies (CNBs), but little clinical validation and a lack of clinical deployment of such tools. We report here on a blinded clinical validation study and deployment of an artificial intelligence (AI)-based algorithm in a pathology laboratory for routine clinical use to aid prostate diagnosis.An AI-based algorithm was developed using haematoxylin and eosin (HE)-stained slides of prostate CNBs digitised with a Philips scanner, which were divided into training (1 357 480 image patches from 549 HE-stained slides) and internal test (2501 HE-stained slides) datasets. The algorithm provided slide-level scores for probability of cancer, Gleason score 7-10 (vs Gleason score 6 or atypical small acinar proliferation [ASAP]), Gleason pattern 5, and perineural invasion and calculation of cancer percentage present in CNB material. The algorithm was subsequently validated on an external dataset of 100 consecutive cases (1627 HE-stained slides) digitised on an Aperio AT2 scanner. In addition, the AI tool was implemented in a pathology laboratory within routine clinical workflow as a second read system to review all prostate CNBs. Algorithm performance was assessed with area under the receiver operating characteristic curve (AUC), specificity, and sensitivity, as well as Pearson's correlation coefficient (Pearson's r) for cancer percentage.The algorithm achieved an AUC of 0·997 (95% CI 0·995 to 0·998) for cancer detection in the internal test set and 0·991 (0·979 to 1·00) in the external validation set. The AUC for distinguishing between a low-grade (Gleason score 6 or ASAP) and high-grade (Gleason score 7-10) cancer diagnosis was 0·941 (0·905 to 0·977) and the AUC for detecting Gleason pattern 5 was 0·971 (0·943 to 0·998) in the external validation set. Cancer percentage calculated by pathologists and the algorithm showed good agreement (r=0·882, 95% CI 0·834 to 0·915; p0·0001) with a mean bias of -4·14% (-6·36 to -1·91). The algorithm achieved an AUC of 0·957 (0·930 to 0·985) for perineural invasion. In routine practice, the algorithm was used to assess 11 429 HE-stained slides pertaining to 941 cases leading to 90 Gleason score 7-10 alerts and 560 cancer alerts. 51 (9%) cancer alerts led to additional cuts or stains being ordered, two (4%) of which led to a third opinion request. We report on the first case of missed cancer that was detected by the algorithm.This study reports the successful development, external clinical validation, and deployment in clinical practice of an AI-based algorithm to accurately detect, grade, and evaluate clinically relevant findings in digitised slides of prostate CNBs.Ibex Medical Analytics.

Sonographic image density represents tissue echogenicity, or the acoustic nature of scanned tissue. It has the potential to distinguish normal from abnormal tissue. Humans are poor assessors of tissue echogenicity, yet their abilities have not been quantitatively compared to more objective methods. We compared the accuracy of image density differentiation by the human eye and by densitometry in the fetal liver model. A group of 60 observers was asked to compare the echogenicity of three pairs of images of a fetal liver. Twenty-three repeated the comparison several weeks later. Image density was measured by electro optical densitometry. The two-tailed signed rank test was used to compare the observer's perceptions with the density values and the repeated observations with the original responses. The density of each fetal liver image in pair A was the same. Image B1 had a 12.3% lower density than image B2, and image C1 had a 6.3% greater density than image C2. Observers were accurate more often when comparing images with large density differences (pair B) than they were when comparing images with the same density (pair A) or with small density differences (pair C). Accuracy was not related to being medically qualified or from an obstetric or gynecological department. Intraobserver variation was not significant. We conclude that the human eye is extremely inaccurate at discerning differences in echogenicity when images have small or no differences in optical density. Densitometry is, therefore, indispensable for accurate image density assessment. Copyright © 1997 International Society of Ultrasound in Obstetrics and Gynecology

Sonographic image density represents tissue echogenicity, or the acoustic nature of scanned tissue. It has the potential to distinguish normal from abnormal tissue. Humans are poor assessors of tissue echogenicity, yet their abilities have not been quantitatively compared to more objective methods. We compared the accuracy of image density differentiation by the human eye and by densitometry in the fetal liver model. A group of 60 observers was asked to compare the echogenicity of three pairs of images of a fetal liver. Twenty-three repeated the comparison several weeks later. Image density was measured by electro-optical densitometry. The two-tailed signed rank test was used to compare the observer's perceptions with the density values and the repeated observations with the original responses. The density of each fetal liver image in pair A was the same. Image B1 had a 12.3% lower density than image B2, and image C1 had a 6.3% greater density than image C2. Observers were accurate more often when comparing images with large density differences (pair B) than they were when comparing images with the same density (pair A) or with small density differences (pair C). Accuracy was not related to being medically qualified or from an obstetric or gynecological department. Intraobserver variation was not significant. We conclude that the human eye is extremely inaccurate at discerning differences in echogenicity when images have small or no differences in optical density. Densitometry is, therefore, indispensable for accurate image density assessment.