A Discriminative Level Set Method with Deep Supervision for Breast Tumor Segmentation

Abstract

Background: Breast tumor segmentation in B-mode ultrasound imaging is important for analyzing, identifying, and diagnosing tumors. The level set is an approach most widely used in breast segmentation, and the refinement is still in progress. However, its effectiveness is harmed by a dearth of semantic information. On the other hand, deep networks contain rich semantic information but loss much influential low-level details.

Method: This paper proposes a novel deep-feature embedded level set group to exploit semantically enriched features for breast tumor segmentation. First, a UNet-based network is trained to extract different features at different stages. Each stage has unique features depiction. Then, a novel level-set method is integrated at the end of each stage to approach more accurate and precise features maps. A new feature-discriminator is devised in the energy function of the level set method to refine the low confidence pixels at the boundaries. Lastly, the outputs of the level set method at different stages are incorporated into final feature maps to further empower the segmentation process. Two datasets comprising 349 breast ultrasound images from various hospitals have been utilized to assess the proposed approach's performance. The model's effectiveness is estimated on different metrics, including Accuracy, Sensitivity or True Positive rate, Specificity or True Negative rate, False Positive rate Dice, and IoU values for both datasets. Furthermore, the efficiency of the model is investigated by performing a comparison with several state-of-the-art classic segmentation methods and deep learning methods.

Result: The proposed method outperformed segmenting breast ultrasound tumors in terms of Dice and IoU for datasets A and B (with p-value < 0.005 against compared methods). Additionally, the performance of the proposed approach is evaluated using the Area Under Receiver Operating Characteristics curve (AUC) and Mean Absolute Error (MAE). Our findings indicate that the proposed method seems to gain superiority over other methods by obtaining a lower MAE rate with the highest value of the AUC.

Conclusion: Experiments determine that our method has obtained the best cut-off to deal with the noticeable glitches present in other approaches and generates more accurate segmentation results for tumors in complex images. Hence, the results confirm the proposed method's effectiveness compared to classic segmentation methods over ultrasound images with blurry boundaries, noise, and intensity inhomogeneity. Moreover, our approach gives unprecedented prediction accuracy and similarity for malignant tumors.

Keywords: Deep learning; Discriminative information; Level set; Medical image segmentation.