Segmentation of skin layers on HFUS images using the attention mechanism

Background and Objective:

The fast development of imaging techniques in recent years has opened new diagnostic paths also in dermatology, where high-frequency ultrasound (HFUS) enables the visualization of superficial structures. At the same time, automated ultrasound image analysis algorithms have started to be widely described in the literature. Although the newest deep learning models can classify the images without the previous segmentation steps, they are often the first part of a computer-aided diagnosis framework that helps further measurements. For the clinical evaluation, the parameters of skin layers: entry echo, SLEB and dermis, are the most important for differential diagnosis and accurate evaluation of treatment process.

Methods:

The paper presents a novel neural network model combining contextual feature pyramid blocks with attention gates to segment skin layers accurately. In addition, a sequential model was tested that pre-segmented the entry echo layer as the most characteristic element in the skin ultrasound image. For the first time, we segmented three skin layers: the entry echo layer, SLEB, and dermis. The developed method is verified using two different HFUS image databases containing images acquired with different ultrasound machines and ultrasound probe frequencies. Measures of models’ performance were proposed, assessing the percentage of cases where the model classified the whole image as background and two focusing on the SLEB layer: percentages of false positive and false negatives detections.

Results:

The average Dice indexes, obtained on the dataset recorded for this study, were 0.95, 0.85 and 0.93, respectively for the entry echo, SLEB and dermis. For models trained without transfer learning, proposed architectures were the only ones that detected the skin correctly every time. Both models achieved the lowest false positive (0.35% and 0%) and false negative (4.48% and 3.66%) rates during the experiments.

Conclusion:

Contextual feature pyramid modules and attention gates allow more accurate detection and segmentation of skin layers. The results obtained are compared with other models described in the literature as efficient for HFUS image analysis, and low false positive and false negative rates speak in favor of our approach.