Deep Learning From Limited Training Data: Novel Segmentation and Ensemble Algorithms Applied to Automatic Melanoma Diagnosis

Abstract

Deep learning algorithms often require thousands of training instances to generalize well. The presented research demonstrates a novel algorithm, Predict-Evaluate-Correct K-fold (PECK), that trains ensembles to learn well from limited data. The PECK algorithm is used to train a deep ensemble on 153 non-dermoscopic lesion images, significantly outperforming prior publications and state-of-the-art methods trained and evaluated on the same dataset. The PECK algorithm merges deep convolutional neural networks with support vector machine and random forest classifiers to achieve an introspective learning method. Where the ensemble is organized hierarchically, deeper layers are provided not only more training folds, but also the predictions of previous layers. Subsequent classifiers then learn and correct the previous layer errors by training on the original data with injected predictions for new data folds. In addition to the PECK algorithm, a novel segmentation algorithm, Synthesis and Convergence of Intermediate Decaying Omnigradients (SCIDOG), is developed to accurately detect lesion contours in non-dermoscopic images, even in the presence of significant noise, hair, and fuzzy lesion boundaries. As SCIDOG is a non-learning algorithm, it is unhindered by data quantity limitations. The automatic and precise segmentations that SCIDOG produces allows for the extraction of 1,812 lesion features that quantify shape, color and texture. These morphological features are used in conjunction with convolutional neural network predictions for training the PECK ensemble. The combination of SCIDOG and PECK algorithms are used to diagnose melanomas and benign nevi through automatic digital image analysis on the MED-NODE dataset. Evaluated using 10-fold cross validation, the proposed methods achieve significantly increased diagnostic capability over the best prior methods.