Abstract:
Surgical excision followed by margin assessment is the most preferable and convenient monitoring
for squamous cell carcinoma (SCC) of skin cancer until tumor-free surgical margins. Nowadays,
SCC margin assessment is done by visual inspection of histopathology images using a
conventional microscope, which is time-consuming, depends on experts’ experience and excellent
visual perception, which may lead to misdiagnosis and mistreatment plans. Different Artificial
intelligence-based computer-aided diagnosis systems have been proposed in the literature to
diagnose margin status from different imaging modalities. However, most of them are based on
the availability of sophisticated imaging devices, their cost limits affordability in a low-resource
setting, and their accuracy needs to be improved. In this study, an automated system is proposed
for the margin classification of SCC to support the manual assessment method by applying deep
learning techniques. The system has been trained, validated, and tested on 828 histopathology
images from Jimma University Medical Center Pathology department from 7 skin sites histologic
grades. The acquired images were preprocessed using median filter to enhance image quality, and
stain normalization to decrease different stain concentrations. In addition to this, rotating at
90°,180°, and 270°, flipped horizontally and vertically were used to increase the size of the
training data set. The overall testing accuracy of the models became 95.2%, 91.5%, 87%, and
85.5% using pre-trained models by fine-tuning the hyper-parameter of EffecientNetB0,
MobileNetv2, ResNet50, VGG 16 respectively. EfficientNetB0 showed the best testing accuracy.
The proposed system helps as a decision support system in a low-resource setting where both
expert pathologists and the means are limited. It helps to assist the healthcare providers in rural
and low-resource areas equipped with less experienced medical experts. Moreover, it helps to
improve surgical outcomes, increase survival rate, and decrease recurrence and reconstruction
time for a better quality of patient’s life. Even though the proposed work achieved the best
performance, further improvement is required by expanding the size of the dataset, including other
common skin cancers
