gms | German Medical Science

Introduction: Automated medical image analysis is an active and highly demanded research field, especially for clinical decision support. One technique of this analysis is medical image segmentation. In this field certain structures in an image are highlighted. Neural networks represent a field of research that can quickly achieve accurate results in the medical image segmentation using a GPU. On the other hand, random forests (RFs) are widely popular in the scientific field of machine learning. They can be trained quickly without the need for a GPU. By implementation as well as application, we compared the performance and usability differences between two random forest pipelines and one deep convolutional neural network model.

Methods: For this study, a random forest whole image architecture (RF-WI) was implemented, which utilized whole images as input by simply defining all pixels as a single feature array. Furthermore, the random forest feature extraction architecture (RF-FE) was implemented, which uses a pixel-wise approach. Therefore, four features per pixel were calculated. These included the pixel-value of the original value intensity, the mean of the 13x13 pixel neighborhood of the pixel in the original as well as of the Sobel filtered image. Additionally, a square of 13x13 pixels of the neighborhood of the pixel of the original image was added to the feature set. To compare the architectures with the cutting-edge, these two random forest approaches were compared with a State-of-the-art deep convolutional neural network (DCNN). This DCNN was built using MIScnn [1]. We performed our experiments on the cell tracking dataset PhC-C2DH-U373 from Dr. Kumar et al. [2] and the retinal dataset “Drive“ from Hoover et al. and a Netherlands screening program [3].

Results: One the PhC-C2DH-U373 the RF-WI achieved a macro-averaged accuracy of 0.95, a Dice similarity coefficient (Dice) of 0.23, an intersection-over-union (IoU.) of 0.16 and a sensitivity of 0.17. The RF-FE achieved an accuracy of 0.98, a Dice of 0.85, an IoU. of 0.75 and a sensitivity of 0.84 with the same dataset. The DCNN model obtained a macro-averaged accuracy of 0.99, a Dice of 0.90, an IoU. of 0.84 and a sensitivity of 0.89 on the PhC-C2DH-U373 dataset.

A macro-averaged accuracy of 0.91, a Dice, an IoU. and a sensitivity of 0.00 were obtained on the retinal imaging dataset with the RF-WI. On this dataset, the RF-FE achieved an macro-averaged accuracy of 0.95, a Dice of 0.68, an IoU. of 0.52 and a sensitivity of 0.58. On the retinal imaging dataset a macro-averaged accuracy of 0.96, a Dice of 0.77, an IoU. of 0.63 and a sensitivity of 0.75 was achieved with the DCNN.

Discussion: The evaluation showed that the DCNN achieved the best results. The whole image architecture, on the other hand, achieved only poor results. However, the RF-FE achieved similar high performance as the DCNN. As disadvantage, we identified the large hardware memory requirement of the RF approaches.

Conclusion: Our results indicate that our RF-FE approach is a performance-wise equal alternative to DCNNs for medical image segmentation without the requirement of a GPU.

The manuscript has already been published on a pre-print server (https://arxiv.org/) [4], which was clarified in advance in consultation with the organising committee of GMDS-TMF-2021.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.