JMIR Publications

ABSTRACT

Background:

Breast cancer is the most common cancer in women globally and mammograms are a primary method for diagnosing it. The challenge of interpreting mammograms is exacerbated by shortage of breast radiologists and lengthy training required.

Objective:

This study aims to evaluate the performance of AI assistance in mammographic reading.

Methods:

Methods:

A multi-reader multi-case study was conducted at the National University Hospital, Singapore from May to August 2023. De-identified mammograms of 500 women were included. Seventeen radiologists read mammograms for six weeks without and with AI assistance, with one-month washout period in between. FxMammo AI software provided heatmaps over areas of suspicion and malignancy risk scores. Diagnostic performance between resident and consultant radiologists were compared with and without AI assistance. Cost analysis was performed.

Results:

Of the 500 women included, 250 had breast cancer whilst 250 had normal or benign breast lesions. Majority had dense breasts (67·4%). The diagnostic performance of AI was significantly higher than consultant radiologists with AUROC of 0·93 (95% CI 0·91-0·95) versus 0·90 (95% CI 0·89-0·92), (p=0·049). With AI assistance, both junior and senior residents showed improvement in diagnostic performance, AUROC from 0·84 to 0·86, p=0·003), and AUROC 0·85 to 0·88, p<0·001) respectively. With AI assistance, AUROC of senior residents was comparable to consultant radiologists with difference in AUROC of 0·02 (95% CI 0·00 – 0·039, p=0·051). Time savings were observed with potential cost savings of SGD$1,361,092 – SGD$1,697,118 per annum.

Conclusions:

Conclusion: This study outlines a cost-saving workflow that utilizes AI assistance to improve diagnostic performance and efficiency of resident radiologists. Clinical Trial: na