Abstract:
This study focuses on the classification of five lethal lung diseases along with
normal lung conditions,COVID-19, pneumonia,TB,CLDs and ILDs. In collaboration
with two doctors, a real-time dataset is gathered, using HRCT
scans to identify the most revealing pathological characteristics of each disease.
Based on the few-shot learning technique, a novel model is proposed
and compared to established architectures such as MobileNetv3Small, MobileNetv3Large,
ResNet18, and EfficientNet. Training the models from scratch
reveals that our suggested model outperforms the other models with 99.47%
accuracy, outperforming them (48%, 50%, 48.3%, and 49% accuracy, respectively).
Furthermore, when implementing transfer learning to pretrain
networks on ImageNet, the other models show promising results when utilized
for few-shot learning. Due to overlapping effects and anomalies across
different diseases, the study emphasizes the difficulties in precisely determining
the root cause of lung ailments using HRCT scans. To address this, a
real-world dataset is gathered to aid the research. The findings highlight the
promising potential of few-shot learning techniques as well as the no need of
large datasets for effective lung disease diagnosis and
categorization.
