MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Benchmark for Breast Ultrasound Image Computer-Aided Diagnosis

E Zhang1*, J Li2 , S Seiler3 , M Chen4 , W Lu5 , X Gu6 , (1) UT Southwestern Medical Center, Dallas, TX, (2) Guangdong General Hospital, Guangzhou, China, (3) UT Southwestern Medical Center, Dallas, TX, (4) UT Southwestern Medical Center, Dallas, TX, (5) UT Southwestern Medical Center, Dallas, TX, (6) UT Southwestern Medical Center, Dallas, TX

Presentations

(Sunday, 7/14/2019) 1:00 PM - 2:00 PM

Room: 225BCD

Purpose: To objectively evaluate the performance of automated breast cancer classification methods, we will publish a B-mode breast ultrasound image classification benchmark (BUSICB) with 800 images and quantify the performance of existing state-of-the-art classification methods.

Methods: The BUSICB dataset with 800 images was retrospectively collected at the University of Texas Southwestern Medical Center, under the approved IRB protocol and compliance to its guideline and regulation. All images were acquired with a Philips iU22 scanner (Philips Medical Systems, equipped with a 12-5 MHz linear array transducer). The average size of the images is 870*660 pixels, with a varied pixel size from 0.04 to 0.1mm (average pixel size is 0.068mm). The dataset comprised 550 benign and 250 malignant lesions, which were associated with a pathologic examination of a subsequent biopsy, with benign and malignant clinically identified. Each image was marked with two or four boundary points by the clinicians during clinical examination. Based on these landmarks, an experienced radiologist delineated each lesion boundary as the reference target regions of interest. We compared and analyzed the performance of the traditional machine learning (Support vector machine with textural and morphologic features) and deep learning methods (deep stacked convolutional auto-encoder classifiers with original image and BI-RADS features) for breast cancer classification in terms of accuracy, sensitivity, specificity and AUC (area under the curve).

Results: A subset of BUSICB dataset including 295 images (205 benign lesions and 90 malignant lesions) have been used in our recent study, in which we proposed BI-RADS features oriented semi-supervised deep learning diagnosis method and achieved classification accuracy around 83.90±3.81% on this subset dataset.

Conclusion: To the authors’ best knowledge, BUSICB will be the largest public dataset for breast ultrasound image classification. This benchmark dataset will help researchers compare their methods with other work objectively and improving their method’s performance.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by the Seed Grant from Department of Radiation Oncology at University of Texas Southwestern Medical Center.

Keywords

Breast, Ultrasonics, CAD

Taxonomy

IM- Ultrasound : Machine learning, computer vision

Contact Email