The challenges in segmentation of HCC lesions from CT scans

The most recent epidemiological data suggest that liver cancer is in the top 10 for both prevalence and mortality (the 6th leading cause of cancer and the 4th most common cause of death) [1]. There is also evidence that its incidence in the?US?and other developing countries is increasing due to an increase in hepatitis C virus infections [2]. Therefore it is only natural that it gained a lot of attention also from the community of machine learning researchers. The most typical type of liver cancer is HCC – Hepatocellular Carcinoma - primary?liver cancer?in adults that is currently the most common cause of death in people diagnosed with?cirrhosis [2].

The detection of HCC tumors is a fairly challenging task. While eventually they grow large, when it happens it is usually a little late. The problem is that it is rarely one HCC tumor. In fact, according to the study published in Hepatobiliary Surgery and Nutrition [3], 75% of HCCs are diagnosed as multiple tumors. Many of them may start to develop simultaneously while still being small and are only detected when the stage of illness is very serious. Therefore, the real problem in diagnosing HCC is diagnosing it early enough, and finding all the small blobs in the image that correspond to cancerous cells.

Fig 1: Example of multiple HCC lesions on liver scan [4]

In the context of machine learning, it also introduces a wide array of problems regarding training and evaluating the performance of the models.

• The most obvious one is the quality of ground truth. The problem is very challenging and even experienced raters may miss some of the tumors.
• Therefore, if that is possible, one should try to mitigate the inconsequence of annotation of small tumors that may confuse the model. It may be tackled by designing a rigorous annotation process, in which multiple raters delineate tumors on the same scans and then jointly work on the common version if the masks they produced differ visibly. The result may be further assessed by the field expert to confirm that masks are in fact correct or need a correction.
• Quite often the authors of the papers that tackle the problem of HCC segmentation mention using only one modality in their work. Usually, these are scans in the portal-venous phase. However, the clinical literature suggests that different features of HCC are visible in different phases of contrast enhanced CT, and usually at least the arterial phase is required additionally to correctly assess some tumors [9] [10]. Some papers use a multiphase CT [5] [6] but they seem to be the minority. Popular datasets in the public domain also provide only one modality, an example being Medical Segmentation Decathlon [7].
• Therefore to create a good model it should be preferable to use both phases as separate image channels for the input to the model if they are available. However, to use these two phases, they have to be coregistered. Coregistration is difficult enough when one lesion has to be well aligned, but in the case of possibly even tens of lesions - this task becomes very hard to accomplish.
• Because the tumors are small and the liver is a large organ, even perfect ground truth will suffer from class imbalance problem, being a problem of a significant disproportion between the majority and minority class examples (here, the liver and HCC voxels belonging to the ground truth). Because of that, it is necessary to account for that while training the model by incorporating loss weighting and other methods [11], like downsampling the dataset by removing all or some slices/patches without the lesions [8].
• Last but not least - performance evaluation. While segmentation quality is usually well measured with the DICE coefficient, it does not tell much about the detection performance. One well-segmented larger tumor may yield a very high overall Dice score even if multiple small tumors were not detected. It is then very important to come up with robust metrics that will be used to measure precisely the quality of detection, not only segmentation.

?Recently we had a chance to address these problems ourselves while working on our HCC segmentation project. We write about it more on our website here: Liver tumor project - Graylight.

Check it out if you want to learn more. Creating a solution for automatic segmentation of HCC is a challenging task. It proves that, above all, data quality and performance evaluation are among the most important parts of a working machine learning system. Ensuring the quality of the ground truth and then carefully comparing to it the result obtained with the trained model is key to reaching the maximum potential of the solution.

?Recommendation by Damian Kucharski

?

[1] “Liver (Hepatocellular) Cancer Prevention (PDQ?)–Health Professional Version - NCI.” PdqCancerInfoSummary, May 23, 2005. Nciglobal,ncienterprise. Liver (Hepatocellular) Cancer Prevention (PDQ?)–Health Professional Version .

[2] “Hepatocellular Carcinoma - The Lancet.” Accessed May 11, 2022. https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(18)30010-2/fulltext .

[3] Baffy, Gy?rgy. “Decoding Multifocal Hepatocellular Carcinoma: An Opportune Pursuit.” Hepatobiliary Surgery and Nutrition 4, no. 3 (June 2015): 206–10.

Decoding multifocal hepatocellular carcinoma: an opportune pursuit .

[4] “Liver Cancer - UCLA Interventional Oncology - Los Angeles, CA.” Accessed May 11, 2022.

Liver Cancer - UCLA Interventional Oncology - Los Angeles, CA .

[5] Ouhmich, Farid, Vincent Agnus, Vincent Noblet, Fabrice Heitz, and Patrick Pessaux. “Liver Tissue Segmentation in Multiphase CT Scans Using Cascaded Convolutional Neural Networks.” International Journal of Computer Assisted Radiology and Surgery 14, no. 8 (August 1, 2019): 1275–84. https://doi.org/10.1007/s11548-019-01989-z.

[6] Kim, Dong Wook, Gaeun Lee, So Yeon Kim, Geunhwi Ahn, June-Goo Lee, Seung Soo Lee, Kyung Won Kim, Seong Ho Park, Yoon Jin Lee, and Namkug Kim. “Deep Learning–Based Algorithm to Detect Primary Hepatic Malignancy in Multiphase CT of Patients at High Risk for HCC.” European Radiology 31, no. 9 (September 1, 2021): 7047–57. https://doi.org/10.1007/s00330-021-07803-2.

[7] “[2106.05735] The Medical Segmentation Decathlon.” Accessed May 12, 2022. The Medical Segmentation Decathlon .

[8] Ayalew, Yodit Abebe, Kinde Anlay Fante, and Mohammed Aliy Mohammed. “Modified U-Net for Liver Cancer Segmentation from Computed Tomography Images with a New Class Balancing Method.” BMC Biomedical Engineering 3, no. 1 (March 1, 2021): 4. Modified U-Net for liver cancer segmentation from computed tomography images with a new class balancing method - BMC Biomedical Engineering.

[9] Hennedige, Tiffany, and Sudhakar Kundapur Venkatesh. “Imaging of Hepatocellular Carcinoma: Diagnosis, Staging and Treatment Monitoring.” Cancer Imaging 12, no. 3 (February 8, 2013): 530–47.

NCBI - WWW Error Blocked Diagnostic .

[10] Nowicki, Tomasz K., Karolina Markiet, and Edyta Szurowska. “Diagnostic Imaging of Hepatocellular Carcinoma - A Pictorial Essay.” Current Medical Imaging Reviews 13, no. 2 (May 2017): 140–53. Diagnostic Imaging of Hepatocellular Carcinoma - A Pictorial Essay .

[11] Krawczyk, Bartosz. “Learning from Imbalanced Data: Open Challenges and Future Directions.” Progress in Artificial Intelligence 5, no. 4 (November 1, 2016): 221–32. https://doi.org/10.1007/s13748-016-0094-0 .