AI Based Diagnosis & Classification Of Lymph Node Fine Needle Aspiration Cytology (FNAC)

In clinical practice, an essential technique that is used by pathologists to diagnose lymph node malignancies is the visual examination of cytopathological slides. Manual visual inspection of whole slide images is challenging, time-consuming, and subject to substantial inter-observer variability, which might result in sub-optimal detection of said malignancies. These False-positive and False-negatives can lead to further complications. Furthermore, upon understanding the current situation, we find that pathologists must review hundreds of slides on an everyday basis. Most of the reviewed slides are normal in nature while only a fraction being abnormal or malignant. The abnormal slides need to be submitted for further analysis. This makes it a very time- consuming and tedious process. Our research looks at pathology from a computational perspective. We leverage deep learning technology to learn features, patterns, and detect anomalies to aid in the detection of lymph node malignancies in whole slide images of fine needle aspiration cytology. An initial screening to classify reactive, infectious, and neoplastic slides would reduce the workload of a pathologist largely. Our main goal was to develop a dataset containing digitized images of reactive, infectious, and neoplastic FNAC slides of lymph nodes. Furthermore, we used this dataset to identify techniques for pre-processing these images such as clustering, thresholding and morphological operations. We then identified and implemented 3 classification models that were appropriate for our dataset, namely, Convolutional Neural Network (CNN), InceptionNet and Resnet50. On evaluation of the results obtained, we found that in terms of accuracy, Resnet50 performed the best amongst all three models with a training accuracy of 76.19%. However, further future improvements such as obtaining a larger dataset and performing feature extraction are necessary in order to increase the accuracy of the models.