JMIR Publications

ABSTRACT

Background:

Given the highly biological heterogeneity of breast cancer, a more effective non-invasive diagnostic tool that unravels microscopic histopathology pattern is of urgent need.

Objective:

This study aims to identify cancerous regions in ultrasound images of breast cancer via convolutional neural network (CNN), using biopsy whole slide images (WSIs) as the reference standard.

Methods:

This single-center study prospectively included participants undergoing ultrasound-guided core needle biopsy (CNB) procedures and pathologically confirmed with breast cancer from July 2022 to February 2023 consecutively. After CNB procedures, the stained breast tissue specimens were sliced and co-registered with ultrasound image of needle tract. CNN models for identifying cancer cells in breast cancer in ultrasound images were developed using FCN-101 and DeepLabV3 networks. The predictive performance was evaluated and compared quantitatively by pixel accuracy (PA), and dice similarity coefficient (DSC). Cancerous region in the testing dataset was further illustrated in ultrasound images.

Results:

A total of 105 participants with 386 breast cancer ultrasound images were included, with 270, 78 and 38 images in the training, validation and test dataset, respectively. Both models performed well whereas the FCN-101 model was superior to the DeepLabV3 model in terms of PA (86.91% vs 69.55%, P = .02) and DSC (77.47% vs 69.90%, P = .045). Furthermore, FCN-101 model had an advantage in predicting cancerous regions, whilst DeepLabV3 model achieved more accurate predictive pixels in normal tissue (both P < .05).

Conclusions:

Breast cancer regions were accurately identified and localized on a pixel level in high-frequency ultrasound images via an advanced convolutional neural network with histopathologic whole slide image as reference standard. Clinical Trial: Ethics [2023]842