Page 90 - Big Data Analytics for Intelligent Healthcare Management

4.8 CONCLUSION 83

Xception+KNN
Xception+SVM
Xception+LR
IRV2+KNN

IRV2+SVM
IRV2+LR Recall
IV3+KNN Precision
Accuracy
IV3+SVM
IV3+LR

RN50+KNN
RN50+SVM
RN50+LR

75 80 85 90 95 100
FIG. 4.30
Test performance graph for 400 .

4.7 DISCUSSION
In this work, several pretrained deep learning architectures were applied for feature extraction instead
of using them as a classifier, which enabled us to save on the time for training. The dimensions of fea-
tures were reduced so that the classifiers could fit them properly within a shorter period of time. The
results reveal that the best validation and test accuracy for each of the magnification factors was quite
impressive. Beside accuracy, we also analyzed the performance in terms of precision and recall since
precision and recall are important for medical image classification as we always want to classify the
tumorous image correctly rather than classifying the nontumorous image correctly and this can be con-
sidered as a tradeoff between precision and recall. The result for each of the combinations of feature
extractor and classifier were shown and interpreted graphically. As it is not wise to make a decision in
medical diagnosis based on a machine learning model, this model can assist the pathologists for
diagnosis of breast tumor detection.

4.8 CONCLUSION
In this work, we have classified breast cancer histopathological images into two major classes—benign
and malignant by our proposed model using some deep feature extractors and supervised classifiers.
The field of machine learning is huge and there are lots of feature extractors and classifiers that can be
used to automate this task. Since the overall performance of this model is not 100%, there is room for
improvement.