Classification of breast cancer in ultrasound imaging using a generic deep learning analysis software: a pilot study

Abstract

Objective: To train a generic deep learning software (DLS) to classify breast cancer on ultrasound images and to compare its performance to human readers with variable breast imaging experience.

Methods: In this retrospective study, all breast ultrasound examinations from January 1, 2014 to December 31, 2014 at our institution were reviewed. Patients with post-surgical scars, initially indeterminate, or malignant lesions with histological diagnoses or 2-year follow-up were included. The DLS was trained with 70% of the images, and the remaining 30% were used to validate the performance. Three readers with variable expertise also evaluated the validation set (radiologist, resident, medical student). Diagnostic accuracy was assessed with a receiver operating characteristic analysis.

Results: 82 patients with malignant and 550 with benign lesions were included. Time needed for training was 7 min (DLS). Evaluation time for the test data set were 3.7 s (DLS) and 28, 22 and 25 min for human readers (decreasing experience). Receiver operating characteristic analysis revealed non-significant differences (p-values 0.45-0.47) in the area under the curve of 0.84 (DLS), 0.88 (experienced and intermediate readers) and 0.79 (inexperienced reader).

Conclusion: DLS may aid diagnosing cancer on breast ultrasound images with an accuracy comparable to radiologists, and learns better and faster than a human reader with no prior experience. Further clinical trials with dedicated algorithms are warranted. Advances in knowledge: DLS can be trained classify cancer on breast ultrasound images high accuracy even with comparably few training cases. The fast evaluation speed makes real-time image analysis feasible.

Figure 1.

Flowchart of the patient selection process.

Figure 2.

Receiver operating characteristic curve of the whole study population (black solid) as well as the test data set (black dashed) and the performance of the human readers on the test cohort (red and orange for the radiologists, purple for the medical student). AUC for the software were 0.96 for the training set, 0.84 for the validation set, and for the readers (validation only) 0.89, 0.89 and 0.79, respectively. AUC, area under the receiver operating characteristic curve.

Figure 3.

A 71-year-old female with scar after segmentectomy of the right breast. This scar was originally classified as BI-RADS 3, before it was down-staged after stable follow-up (total follow-up: 28 months). Both the neural network (0.95, cut-off 0.69) and the two radiologists (4/5 and 5/5) rated the lesion false positive as probably malignant. BI-RADS, Breast Imaging Reporting and Data System.

Figure 4.

One of the rare examples of the false negatives, where the human readers were superior to the neural network in detecting malignancy. A 41-year-old female with a palpable mass in her left breast initially rated as BI-RADS 5 lesion and later confirmed malignant (invasive ductal carcinoma). While the neural network rated the lesion as rather benign (0.46, cut-off 0.69), the two readers with clinical experience classified it as probably malignant (4/5) and the medical student as undetermined (3/5). BI-RADS, Breast Imaging Reporting and Data System.

Figure 5.

A 60-year-old female with an initially BI-RADS 4 classified lesion of the right breast, which turned out to be a cyst after biopsy. The two radiologists rated the lesion the same as the radiologist performing the examination had done, as somewhat between indifferent and rather malignant (3 and 4/5), while the medical student rated the lesion as rather benign (2/5). The neural network classified the lesion correctly as benign and could have prevented the unnecessary biopsy (0.23, cut-off 0.69).

Figure 6.

The only male patient (58 years old) and the only lymphoma included in the study population were rated as rather benign by the two radiologists (2/5), but as malignant by the neural network (0.78, cut-off 0.69). Interestingly, the medical student also correctly diagnosed the lesion as potentially malignant (4/5).

Figure 7.

A 60-year-old female with a lesion initially rated as BI-RADS 4, later confirmed as biopsy-proven fibrosis of the left breast and 28 months of unsuspicious follow-up. All the human readers rated the lesion as probably malignant (4 or 5/5). Only the neural network classified the lesion correctly as benign (0.38, cut-off 0.69). This is one of the examples where the neural network could have prevented an unnecessary biopsy.

Figure 8.

A 55-year old female with an initially BI-RADS 4 classified lesion of the left breast, for which biopsy showed adenosis and no sign of malignancy. It was correctly classified as benign by the neural network (0.52, cut-off 0.69), which might have rendered the biopsy unnecessary.