Abstract
1432
Objectives: One of the most challenges in medical imaging is the lack of data and annoted data. It is proven that classical segmentation methods such as U-NET are useful but still limited due to the lack of annotated data. The aim was i) to devdelop and validate a new weakly supervised learning method to segment lung and oesophageal lesions in FDG PET imagnig, ii) predict patient’s response and survival based on the segmentation results, and iii) to compare its performance with state of the art methods based on supervised learning.
Methods: We have developed a new architecture to segment lung and oesophageal lesions, and then perform an outcome prediction based on a weakly supervised learning approach. We have also evaluated state of the art architecture U-NET (Ronneberger et al. MICCAI 2015) for the segmentation and 3-D RPET-NET (Amyar et al. IEEE TRPMS 2019) for outcome prediction. Using a weakly supervised learning is a promising way to address the lack of annotated data problem, however, it is challenging to train one model to segment efficiently different type of lesions due to the huge variation in images. We present a novel approach to segment different type of lesions in PET scan images in 3D using only a class label at the image level. First, a simple convolutional neural network classifier is trained to predict the type of cancer on two 2D images. Then, a pseudo-segmentation of the tumor is generated using class activation maps, back-propagated and corrected in a multitasking learning approach. Finally, we use the mask generated from the two 2D images to segment the tumor in the 3D image.
Results: All the methods were compared based on the ability to detect accurately the tumor and to conduct an outcome prediction. The performances are measured by dice coefficient for segmentation, accuracy, sensibility, specificity and the area under the ROC curve for the prediction. The results were obtained using a 5 fold cross-validation. Best results for segmentation were obtained using our proposed model for both lung and oesophageal cancer (0.77+-0.07 and 0.73+-0.09 vs 0.63+-0.14 and 0.53+-0.17). For radiomics, 3d-rpet-Net with manual segmentation was not statistically significantly different from our model (p=0.59) for oesophageal and (p=0.63) for lung (AUC= 0.70+-0.04 and 0.61+-0.03 vs 0.67+-0.08 and 0.59+-0.04). Our model tend to have a better sensibility for oesophageal and a better specificity for lung cancer with no significant differences.
Conclusions: A new weakly supervised learning model was developed to localize lung and oesophageal tumors in PET images. It utilizes two fundamental components: a new class activation map to locate the tumor and a new loss function to improve localisation precision. The model could detect tumors with better accuracy compared to fully supervised models such as U-NET, or classical CAMs. Our model outperformed other methods in terms of the dice index. As for radiomics analysis, 3d-rpet-net with manual segmentationis showing slightly better results than our model in radiomicsanalysis. However, it is based in manual pixel-level annotations of tumor, which requires a physician expert and also is time consuming. This simple and yet powerful technique, can be integrated in futureworkflow/software dedicated to automatic analysis of PET ex-ams to conduct radiomics analysis.