Performance of a Transfer Learning Module with Pretrained Neural Network and Logistic Regression in Detection of Breast Carcinoma from Microphotographs of Fine Needle Aspiration Cytology Smears
2 Undergraduate student, Vellore Institute of Technology, Katpadi,Vellore, Tamil Nadu, PIN 632014, India
3 Dept of Pathology, Command Hospital, Alipore, Kolkata 700027, India
Lithium ion batteries (LIBs) are highly efficient. They have high capacities and long cycle Fine needle aspiration cytology (FNAC) is an useful modality for initial assesment of a palpable breast lump. The aim of this study was to develop a computer model for classification of microphotographs from FNAC smears of breast lesions, into two classes ‘benign’ and malignant’. We have used the transfer learning method, i.e. using a neural network which has been trained on a different dataset, to extract features from the present datatset. Apart from being of valuable diagnostic utility, the model will also provide key insights on machine learning and how a learner, human or machine, distinguishes benign from malignant.
A pretrained neural network (VGG16) which has been trained on the ImageNet database, was used for the study. A total of 2037 processed microphotographs from Romanowky stained FNAC smears were taken, all at 40x magnification. Images from two different microphotography systems in two different tertiary care centers of India was used. The images were then split into two sets, ‘training’ (1544 images) and ‘validation’ (493 images). During training, features were extracted with VGG16 and fit with original labels using logistic regression. After completion of training, images from the validation set was processed with the VGG16 network and the trained logistic regression model was used to generate predictions.
The model achieved 90.38% sensitivity, 87.12% specificity, 88.67% positive predictive value and 89.03% negative predictive value. A diagnostic accuracy of 89% was achieved. Receiver operating characteristic shows area under curve of 0.89, indicating good perforamance. 12.8% false positives and 9.6% false negatives were also reported by the model. The principal difficulties encountered were the distinction between the dark staining nuclei of myoepithelial cells and the hyperchromasia of a malignant epithelial cell. Also, hypocellular foci with single malignant epithelial cells were often reported to be falsely negative by the model.
Overall, the sensitivity, specificity, positive and negative predictive value of the model is close to FNAC reported by pathologists. It shows potential to be used as a screening tool, after validation on a larger dataset.
Keywords: Pretrained Neural Network, Logistic Regression, Breast Carcinoma, Microphotographs