METHODOLOGY

Image patches are reshaped into square patches by extending the shorter dimension and subsequently resizing to images of 200 by 200 pixels (see Fig. 1). In total: 3761 mass-containing patches are curated; with class counts of 544 (0), 1730 (1), and 1487 (2).

We employ a ResNet-22 conv-net model with ~11.5M trainable parameters and use a random 80/10/10 train/validation/test split. Models are trained for 150 epochs using a cosine-annealed learning rate schedule.

Data-augmentation development:

Data augmentation (DA) provides a proven means for enhancing model performance and generalisability in situations with scarce data availability (see Fig. 2 for DA examples).However, it remains a challenge to determine relevant augmentation magnitudes for the different operations applied to images on the fly. This is addressed by ControlAugment (Ctrl-A) [3], which uses control theory to dynamically balance the training-validation-loss ratio by sequentially updating the DA magnitudes to realise adaptive regularisation. The procedure for determining the maximum augmentation magnitude, Γ, for an operation is visualised in Fig. 3.

RESULTS

Classification results are summarized in Figures 4 and 5 and in Table 1. Figure 4 demonstrates the adaptable regularization enabled by Ctrl-A and the confusion matrix in Fig. 5 illustrates how virtually all misclassification are caused by overlap with the intermediate “ill-defined” class.

CONCLUSION

We have demonstrated the potential of machine-learning-based computer vision, coupled with control-driven data augmentation, for classifying breast mass types in mammographic images.

REFERENCES