Automated volumetric estimation of six basic thoracic and abdominal organs in postmortem computed tomography data using deep learning techniques

Abstract

Purpose

Computed tomography (CT) has become a widely adopted and standard procedure as an adjunct to autopsies in numerous countries. However, owing to the high number of cases and the limited availability of skilled practitioners, the need to streamline the diagnostic process has spurred the advancement of automated solutions. These solutions leverage deep learning methodologies to potentially automate diagnoses by analyzing postmortem CT data. Here, we show how deep learning techniques enable segmentation and volume evaluation to be concurrently performed for six basic thoracic and abdominal organs in postmortem CT data: the heart, lungs, liver, spleen, kidneys, and urinary bladder. Based on these automated volumetric estimations we automatically derived the weight of the heart, lungs, liver, spleen, and kidneys.

Methods

We developed a convolutional neural network tailored for conducting volumetric data segmentation in postmortem computed tomography images based on the U-Net architecture.

Results

Our best model achieved an overall Dice score (F₁ score) of 0.907±0.029. The heart, lung, and liver yielded higher scores than did the spleen, kidneys, and urinary bladder. We also automated the weight calculation of the heart, lungs, liver, spleen, and kidneys.

Conclusion

Our study demonstrated that a convolution neural network such as U-Net could reliably estimate concurrently the volumes of six basic thoracic and abdominal organs from postmortem CT data. Our study also shows how this information can be subsequently used to automatically estimate their weight. However, post- and perimortem changes pose substantial challenges for automatically processing postmortem CT data.