Synopsis

Keywords: Cancer, Breast, TNBC

Quantitative image analysis of cancers requires accurate tumor segmentation that is often performed manually. In this study, we developed a deep learning model with the self-configurable nnU-Net for automated tumor segmentation on dynamic contrast enhanced MR images of triple negative breast cancers. Our results on an independent testing dataset demonstrated that this nnU-Net based deep learning model can perform automated tumor segmentation with high sensitivity and Dice coefficient.

Introduction:

Triple-negative breast cancer (TNBC) is a biologically aggressive and heterogenous subtype of breast cancer that is treated with neoadjuvant systemic therapy but has varied responses, which predict long-term outcome. Quantitative imaging has been investigated and shown potential to predict the early¹. As an important step in the quantitative image analyses, tumor segmentation is usually performed by specially trained personnel in a manual and often laborious manner. Manual tumor segmentation is time-consuming and susceptible to inter- and intra- reader variability. In recent years, deep learning (DL) models have shown great success and potential for automated biomedical image segmentation applications^2-4. In particular, nnU-Net⁵ is a self-configurable DL framework and its segmentation performance is comparable or even surpasses many customized DL models in terms of Dice coefficient (DICE). In this study, we developed a DL model by using nnU-Net on dynamic contrast-enhanced (DCE)-MR images of TNBC patients acquired at baseline before the initiation of therapy and we evaluated the model segmentation performance when different images were used as the model inputs.

Method:

Datasets: We included 162 biopsy-confirmed stage I-III TNBC patients in our study. All the patients were enrolled in an IRB approved prospective clinical trial (NCT02276443) and underwent baseline DCE MRI scan; 108 and 27 patients were used for training and validation, and another 27 patients for testing. DCE images were acquired using a 3D T1-weighted DISCO sequence axially and with the following parameters: acquisition matrix: 320×320, field-of-view:300×300mm, number of slice: 112~192, slice thickness: 3.2mm, slice gap: -1.6mm, TR/TE=3.9/1.7ms, flip angle= 12°, number of phases: 45~62, temporal resolution: ~10s. The segmentation of tumor ROIs from early phase (60s) DCE subtraction images was carried out by two fellowship-trained breast radiologists. The manually segmented tumor ROIs were used as the reference standards for the model development and testing. All DCE images and masks were zero-padded to 192 slices before model training.
Model Training: We trained 3D nnU-Net with default hyperparameter setups that were provided by the model developers as the public download of the source code⁵. Five-fold cross-validation was used on training/validation sets. Five different combinations of raw DCE images and semi-quantitative parametric maps were used as the model inputs: 1. DIFF: difference images between the DCE images at 60s after contrast agent injection and the pre-injection images, 2. PEI: positive enhancement integral images, 3. SER: signal enhancement ratio images, 4. MSI: maximum slope of increase images, 5. COMB: the combination of DIFF, PEI, and MSI images. The PEI, SER, and MSI images were calculated using the MR scanner vendor provided software. The first 4 inputs were trained in 3D mode and COMB was trained in 4D mode by which DIFF, PEI and MSI images were concatenated as the 4th dimension.
Model Testing: Segmentation performance was evaluated on the independent testing dataset with DICEs, and sensitivity for each subject, then summarized over the testing set⁶.

Result & Discussion:

Models with DIFF, PEI, MSI, and COMB as inputs yielded good segmentation performance with DICEs, and sensitivity > 0.8 (Fig. 1). The model performance in one subject is shown in Fig. 2 (top row). In comparison, the model with SER as input performed poorly, with group DICE/sensitivity = 0.46/0.52. The model with DIFF as input yielded the best segmentation performance with a DICE of 0.82 and sensitivity of 0.92, which is slightly lower than that of COMB (sensitivity of 0.93). In one subject, the model with COMB failed to identify the tumor (sensitivity of 0) while the model with DIFF yielded a sensitivity of 0.71 (subject 2 in the bottom row of Fig. 2). Upon DCE image examination, this subject had a large artifact from the biopsy clip at the center of the tumor, which led to an excluded hollow area within the tumor in the manually contoured tumor ROI. It is conceivable that the current model requires additional cases for training to further improve the model performance.

Conclusion:

We have successfully developed a DL model for automated tumor segmentation using nnU-Net on DCE MR images of TNBC. Our investigation showed that different combinations of the DCE images and semi-quantitative parametric maps can be used as inputs to the model. An optimal model may be established with a larger number of patients for training and testing. Such a model will be highly useful for quantitative image analyses in a variety of clinical applications.

Acknowledgements

No acknowledgement found.