The project was developed as final project for the Computational Intelligence course. The classifiers are built from scratch or pretrained models such vgg16 and incepcionV3. and one awesome ensemble system. the CI_project holds the entire documentation and if you have problems accessing the files, let me know. The Classifiers are as follows:
- Scratch_CNN_2classes design of a from-scratch classifier to discriminate among Masses and Calcifications.
- Scratch_CNN_4classes design of a from scratch classifier to discriminate among benign mass, malignant mass, benign calcification and malignant calcification.
- Pre_Trained_2classes design of a pre-trained models, a comparison between vgg16 and InceptionV3 as full-fine-tune and feature-extractor-only to discriminate among mass and calcification.
- Pre_Trained_2classes design of a pre-trained models, a comparison between vgg16 and InceptionV3 as full-fine-tune and feature-extractor-only to discriminate among benign mass, malignant mass, benign calcification and malignant calcification.
- BaselineCNN design of context-based classifier to discrimate masses and calcification. it compares InceptionV3 as siamesse and full-fine-tuned model.
- Ensemble Several mathematicals models are proposed to merge the output of the following models Scratch_CNN_2classes, Pre_Trained_2classes and, BaselineCNN. Such math models were: Average, Accuracy-based Weighted average, Logic Voting (inspited in Redundancy Engineering), Precision-based Weighted Average and, Opinion Rating.
the idea behind this approach is to ignore as much as possible the net's output around 0.5, and emphasize or keep the values close to 1 and 0, because the last values define the classification. For such purpose, three equations were analysed. tan, sinh, 3-degree polynomial, 5-degree polynomial, 29-degree polynomial and, 63-degree polynomial. The large polynomial values were used to understand how much outputs around 0.5 can be considered meaningless before compromise the good performance.