A Data-Efficient Deep Learning Approach for Lung Nodule Segmentation

*Murtala Dattijo Makama1, Abdulkarim Salihu Aish1 , Ibrahim Dattijo Makama2

1National Information Technology Development Agency (NITDA)
2National Open University of Nigeria (NOUN)

*Corresponding author’s Email: mdattijo@nitda.gov.ng, doi.org/10.55639/607.02010006


ABSTRACT

Medical imaging is undeniably a very important field in the provision of effective health care. It allows for early detection, diagnosis, and treatment of different kinds of diseases. On the other hand, preparation of labeled data from the biomedical domain poses difficulties, since it requires heavy privacy considerations and the expert radiologist to perform manual labeling of that data. Due to that reason, efforts to utilize advancements in deep learning algorithms are often hampered. Generating expert annotation of medical images at scale is difficult, costly, and time-consuming. This challenge is even greater in developing countries, especially in Africa, where there is a massive shortage of radiologists to annotate these images. Another related issue in medical imaging is the scarcity of datasets, particularly in Africa, where there is a severe shortage of digital imaging modalities. This becomes more obvious when studying a particular disease, for instance. Hence, the scarcity of large, well-labelled datasets is the main constraint for deep learning applications in medical imaging. To solve this problem, we devise a self-supervised learning approach. The proposed method used inpainting as the pretext task and U-Net for segmentation. The pre trained network was then fine-tuned on a varied sample of annotated datasets. Our results performance improvement with 50% of labeled data compared to supervise learning. This approach has the potential to significantly improve the applicability of deep learning for medical imaging in resource-constrained settings

Keywords:

Self-supervised
learning,
Deep learning,
Medical imaging,
Segmentation,
Lung CT scans