[ePoster] Predicting axillary lymph node metastasis in early-stage breast cancer using primary tumor image features on [18F] FDG PET: a comparative study of engineered radiomics, deep learning, and conventional methods

Predicting axillary lymph node metastasis in early-stage breast cancer using primary tumor image features on [18F] FDG PET: a comparative study of engineered radiomics, deep learning, and conventional methods

ABSTRACT
Aim/Introduction: Axillary lymph node (ALN) assessment is a key step in breast cancer (BC) management. Yet, the non-invasive evaluation of ALN involvement using imaging lacks sensitivity, and sentinel lymph node excision procedure remains the gold standard. Imaging studies suggested that primary tumor (PT) features might be associated with ALN status _[1-3]. In this context, we investigated the association between ALN status and PT local features characterized using voxel-wise radiomics _[4] and deep learning.

Methods: [18F]FDG PET scans and clinical data from 191 early-stage BC were analyzed. Each PT was labeled according to its ipsilateral ALN status determined after surgery. First, 93 voxel-wise radiomic feature maps per VOI were extracted by sliding a 5×5×5-voxel kernel. The mean value was used for each map to yield a feature vector per tumor. Feature selection and bagging regularized logistic regression were used to build a probabilistic model, M1, to identify positive ALN. Secondly, an analoguous convolutional neural network (CCN) based on the U-net architecture was trained, leading to learned fine-resolution feature maps. VOI-pooling was added before the logistic layer to transform this segmentation network into a classification model, M2. A third model M3 was built based on conventional image features (SUVmax, SUVmean, volume, MTV, TLG, maximum diameter, second diameter, third diameter). For interpretability, radiomic decision maps (RDM) for M1 and class activation maps (CAM) for M2 were computed.

Results: here were 94 histologically positive ALN and 97 negative. Through cross-validation, three features were selected for M1 (GLCM cluster prominence, GLDM LDHGLE, GLDM SDLGLE) and only one for M3 (maximum diameter). The average AUROC was 0.68±0.08 (permutation p-value = 0.01) for M1, 0.65±0.06 (p-value = 0.04) for M2 , and 0.70±0.08 (p-value = 0.01) for M3. By comparing RDMs and CAMs, we found that the voxel-level information extracted by both learning approaches were very similar.

Conclusion: Neither radiomics nor CNN outperformed conventional image analysis (measurement of the largest dimension of the PT) to predict ALN status from PT. Voxel-wise radiomics and CNN captured similar information, without any substantial difference in performance for this ALN status prediction task. Combining imaging data with other modalities such as histology and genomics, as well as clinical information, could allow for improvement.

References: [1] Tseng et al., Med Sci Monit, 2014;20:1155-1161. [2] Zhou et al., Radiology, 2020;294(1):19-28. [3] Huang et al., npj Breast Cancer, 2018;16;4:24. [4] Escobar et al., Med Phys, 2022.