Deep learning with probabilistic models for segmenting lung lobes on computed tomography images with severe abnormalities

Project developed for XXIX Congresso Brasileiro de Engenharia Biomédica (CBEB) 2024 (https://sbeb.org.br/cbeb2024)

Abstract

The development of efficient and robust algorithms in automated tools for segmenting the lung and its lobes is useful for the diagnosis and monitoring of lung diseases that cause lung abnormalities, such as pneumonia caused by COVID-19 and cancerous nodules. The amount of available data containing manual annotations of the lobes in patients with severe lung abnormalities such as consolidations and ground glass opacities is scarce, due to the difficulty of visualizing the lobar fissures. This work aims to develop a method for automated segmentation of lung lobes using deep neural networks in computed tomography images of the lungs of patients with severe abnormalities. The method is based on probabilistic models built from labels fusion, used not only to guide the deep neural networks while learning to segment the lobes, but also for postprocessing the network prediction to obtain the final segmentation. Segmentation is performed in two stages: a coarse stage working on downsampled images and a second high-resolution stage, where specialized AttUNets compete for each lobe's segmentation. The performance of the proposed approach was assessed using two public datasets with lobe annotations, in the presence of cancer nodules and COVID-19 consolidations. Open source implementation is available at

To install dependencies

sh requirements.txt

To run project

python main.py

Project

• This work was submitted to ISBI 2024 (https://biomedicalimaging.org/2024).

• Diagram of the method implemented using CCN and a priori information, incorporated into the network input as probabilistic models:

  

• Segmentations using the LobePrior, nnUnet and LungMask methods on the CT image coronacases_007 (slice 150), from the Coronacases[1] dataset. Here, it is possible to visualize the segmentation errors (rectangular region) generated by each of the networks, in relation to the \textit{gold standard} (Fig. b):

  
  CT image and Ground Truth
  
  
  LobePrior (Dice score = 0.953), nnUnet (Dice score = 0.943) and Lungmask (Dice score = 0.945)
  

• Qualitative evaluation using 3D representations in CT images of lungs from patients with severe injuries:

  

• Qualitative evaluation using 3D representations in CT images of lungs from patients with severe injuries:

  

Citation

@ARTICLE{Jean2024,
	title = {Deep learning with probabilistic models for segmenting lung lobes on computed tomography images with severe abnormalities},
	journal = {CBEB 2024},
	pages = {1-6},
	year = {2024},
	author = {Jean Antonio Ribeiro and Diedre Santos do Carmo and Fabiano Reis and Leticia Rittner}
}