University of Twente Student Theses
Development of a deep learning algorithm to identify mammograms with increased risk of lesion masking.
Willemsen, B. (2024) Development of a deep learning algorithm to identify mammograms with increased risk of lesion masking.
|
PDF
8MB |
| Abstract: | Over the last decades, the Dutch breast cancer screening program has helped reduce mortality of breast cancer. However, a quarter of breast cancers in participating women are not found through this screening process. The main cause of false negative screening results is lesion masking, which occurs when lesions are superimposed on fibroglandular tissue, rendering lesions indiscernible and increasing the risk of them being missed. Identification of breasts with a high risk of lesion masking could be a reason to screen these women with alternative screening methods that are less susceptible to lesion masking. A study by Mainprize et al. has shown that this is possible with a model observer, reaching an area under receiver operating curve (AUC) of 0.79 (95% CI 0.69-0.87) on the original North-American cohort. However, it only reached an AUC of 0.62 (95% CI 0.59-0.65) on a Dutch screening cohort. Therefore, this study focuses on finding an alternative deep learning-based method to predict the risk of lesion masking, that is suitable for the Dutch screening cohort. The main challenge of predicting lesion masking risk in screening mammograms, is the lack of a good ground truth. The majority of participants never get diagnosed with cancer, and for these women it is unknown what their lesion masking risk is. The lesion masking risk can only be approximated for participants that were diagnosed with cancer. In a mammogram with a positive screening result, the lesion masking risk is assumed to be low. When a cancer developed within two years after a negative screening result, the lesion masking risk is assumed to be high. For the dataset of a single screening unit with 76,976 participants, this resulted in a dataset of 1,532 mammograms with low and 426 mammograms with high lesion masking risk. Using only these mammograms, a baseline model with a ResNet18 architecture was trained with five-fold cross validation. The average prediction of two parallel models for both mammographic views resulted in an AUC of 0.62±0.02, which is comparable to the performance of the model observer. For deep learning approaches, the size of the lesion masking risk dataset is limited. Transfer learning is a useful strategy to help a model by making it learn from more data in a similar domain. Volumetric breast density is a value that can be calculated by Volpara® DensityTM (v1.5.4.0, Volpara Health, Wellington, New Zealand). Because this metric holds predictive power towards lesion masking risk, it was chosen as a pre-training task. A ResNet18 and SwinV2 network were developed to predict volumetric breast density with a dataset containing 17,222 negative screening mammograms. Both networks achieved high correlations coefficients of above 0.90. After finetuning, the networks achieved an AUC of 0.63±0.02 and 0.63 ± 0.03 respectively, which are both comparable to the baseline model. An analysis of class activation maps suggests the model partially unlearns density-related features during finetuning. To determine whether predicting volumetric breast density was a suitable pretraining method, an alternative transfer learning method was implemented. A ResNet50 was trained in a different project with a contrastive self-supervised learning method to extract features from mammograms. Extracted feature vectors from the lesion masking risk data set were used to train a small multi-layer perceptron. Here, an AUC of 0.61±0.02 was achieved. The implemented transfer learning strategies failed to yield significant improvement on this baseline, despite good performance in the pre-training task of predicting density. The class activation maps before and after fine-tuning suggest that the model partially unlearns density-related features, suggesting inaccurate labels of lesion masking risk. Finetuning the model on a more restricted dataset showed an improvement to an AUC of 0.70±0.03, which is a large improvement when compared to the baseline model. An analysis of the potential impact of the model shows the possibilities to find otherwise masked lesions by including women in alternative screening methods. |
| Item Type: | Essay (Master) |
| Clients: | Radboud University Medical Center, Nijmegen, Netherlands |
| Faculty: | TNW: Science and Technology |
| Subject: | 44 medicine, 54 computer science |
| Programme: | Biomedical Engineering MSc (66226) |
| Link to this item: | https://purl.utwente.nl/essays/99021 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page