MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Comparative Study with Multiple Losses in Deep Learning for Tumor Segmentation in PET

S Liu1 , X Tan2 , X Zhao1, L Li1 , W Lu3 , S Tan1*, (1) Huazhong University of Science and Technology, Wuhan, Hubei, (2) Hunan University of Technology, Zhuzhou, Hunan, (3) Memorial Sloan Kettering Cancer Center, New York, NY

Presentations

(Sunday, 7/29/2018) 4:30 PM - 5:00 PM

Room: Exhibit Hall | Forum 5

Purpose: Recently, the deep learning methods have been successfully applied into various medical image analysis tasks. Most of these mehods used the cross entropy loss function. To study the influence of different loss functions for tumor segmentation in PET, several losses and their combinations coupled with a 3D fully convolutional neural network (3D FCN) were tested in this study.

Methods: The proposed method learned features from PET images through the supervision of multiple loss functions. The bone architecture of the FCN was 3D U-Net. The loss functions studied included the softmax loss (â„“soft) which is the baseline loss, the SVM loss (â„“svm), the dice loss (â„“dice), and the intra-class feature distance loss (â„“dist) and their combinations. A total of six loss functions were compared. A dataset with 74 non-small cell lung cancer (NSCLC) PET images were used for evaluation, among which 48 images were used as training samples and the remaining 26 images were testing samples. Performance of these losses were evaluated by Dice Similarity Index (DSI), Classification Error (CE) and Volume Error (VE).

Results: The model FCN + â„“soft + â„“dice achieved the highest mean DSI value (0.8632), while the baseline model FCN + â„“soft ranked the last but one and the FCN + â„“dice ranked the last (0.8534). The second to fourth positions were occupied by FCN + â„“svm, FCN + â„“svm + â„“dist, and FCN + â„“svm + â„“dice, respectively. Four out of five models outperformed the basline model, indicating that multiple loss functions were better than the softmax loss for tumor segmentation in PET.

Conclusion: Different loss funcions can affect the performance of deep learning for tumor segmentation in PET. The combination of the dice loss and the softmax loss could significantly improve the segmentation accuracy.

Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by National Natural Science Foundation of China (NNSFC), under Grant Nos. 61375018 and 61672253.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email