Matilda Bilén will present her master's thesis.

Title: Deep Learning for Breast Cancer Diagnosis Combining Mechanical Imaging and Mammography

Author: Matilda Bilén

Supervisor: Anders Heyden, Magnus Dustler, Anna Bjerkén

Examiner: Karl Åström

Abstract: 

Background: Breast cancer is the most common cancer among Swedish women. While the incidence
has increased, the total mortality has decreased significantly, thanks to better treatment as well as
intensive population screening. The screening consists of taking digital mammographies (DMs), which
are assessed by radiologists. If the risk of cancer is deemed increased, the women are recalled
for further assessment. Approximately 80% − 90% of recalled cases are false positives, leading to
unnecessary patient anxiety and increased clinical workload. Mechanical Imaging (MI), measuring
the pressure over the breast during the DM, has been proposed as a supplement to DM to reduce the
number of false positives. This project aims to explore the use of deep learning methods for breast
cancer diagnosis using mechanical imaging (MI) as well as combined mechanical imaging and digital
mammography (DM).
Methods: First, the corresponding images acquired for the two modalities were automatically aligned
using normalized cross-correlation and entropy based image registration. Two separate convolutional
classifiers were then implemented for each modality. A custom CNN was implemented for the MI
data (of sizes 28×20), and for the larger DM images (resized to 448×320), a ResNet18 pretrained on
ImageNet was employed. These models were trained on each modality separately, to later fuse them
in the two multimodal models, which were trained on both modalities together. The multimodal models
were (1) a 1D Fusion Model concatenating the two model outputs at the fully connected layer, and (2)
a 2D Fusion Model concatenating the feature maps (of size 14x10) before the final average pooling
layers, to take advantage of any spatial relationship. Fivefold cross-validation was used to evaluate
the models. The mean AUCs and accuracies were reported.
Results: The mean AUC of the cross-validation was 0.618 for the MI model. The DM model performed
stronger with an AUC of 0.684. The two multimodal models performed similarily to the DM model with
the same mean AUC.
Conclusion: Although the AUC was low (0.618), we can conclude that the MI model could find a
predictive signal from solely the MI data. The DM model performed stronger. There was no evident
improvements for the multimodal models which implies that for these models the signal from the DM
data outweighed the additional signal from the MI data. This could be due to overlapping information
in the modalities, a stronger backbone from the DM model and possibly weaker training due to a
smaller training set for the multimodal models. A stronger performance could be expected with a
larger dataset.