Emerging Semi-Supervised Skin Lesion Segmentation using Unlabelled Samples with GAN and Transformer Representations

Early cancer detection and diagnosis depend on
accurate and automated skin lesion segmentation helping
dermatologists in their work. On the other hand, deep learning
models find great challenge in the absence of large, annotated
datasets in the medical field. These problems consist in
overfitting and limited generalisation. Manual labelling calls for
knowledge, time, and money; hence, depending on very large,
annotated datasets is generally not practical for medical
applications. We introduce a new semi-supervised method using
unlabelled samples including generative adversarial networks
(GANs). This approach is proposed to transcend the already
mentioned limitations. The method makes advantage of a twostage training structure. First phase of the supervised learning
process uses annotated data to learn semantic segmentation
maps. After that, the unsupervised phase uses Structured
Prediction-based Deep Reinforcement Learning (SP-DRL) to
improve generalisation, which lowering the dependence on
annotations. Strong data-driven features have been proposed to
be derived from a surrogate task using convolutional and
Transformer-based representations, which reducing the demand
for annotations. The proposed method was evaluated using three
benchmark datasets: PH2, ISIC 2017, and ISIC 2018. The
segmentation performance shows to be better with increasing
Dice Similarity Coefficient (DSC) scores of 89.2% (ISIC 2018),
87.4% (ISIC 2017), and 91.1% (PH2) respectively. Further, the
model shows better generalisation and resistance against domain
changes, which surpassing conventional approaches under total
supervision. This semi-supervised method closes the difference
between the availability of annotated medical datasets and those
without labels, which enabling the useful adoption of automated
lesion segmentation systems.