As artificial intelligence (AI) technology has become more popular, it has been applied to a variety of fields, such as health care and manufacturing, as well as to everyday life. This technology is powered by advanced computing power, large amounts of data, and new algorithms. By using these tools, the efficiency and productivity of certain processes can be improved.

Accordingly, machine learning (ML), which is a subset of AI, gets a lot of attention these days. ML is a process by which computers learn to identify patterns and correlations in data, without being explicitly programmed to do so. It is often used in conjunction with big data analytics for more accurate results. This process can be used to classify categories, predict future or uncertain conditions, or find relationships in data.1

Another type of advanced computer power is medical image computing (MIC). MIC is a vital field of study that employs computer science and engineering techniques to solve medical imaging problems. MIC’s primary purpose is to extract information or knowledge from medical images that can be used for therapeutic purposes.

Continue Reading

While closely connected to medical imaging, MIC focuses on image computational analysis rather than image capture. This field has many applications in health care, such as diagnostics, treatment planning, and patient care. MIC researchers are constantly developing new methods and tools to improve the way medical images are used.

A subset of ML, deep learning (DL), is one of these improvements for advancing the utilization of medical images. DL uses multiple layers of processing to progressively extract higher-level features from raw input data. In image processing, for example, lower layers may detect edges, while higher layers may identify specific concepts relevant to humans, such as digits, letters, or faces.2

Deep learning algorithms are now used to detect disease or abnormalities from x-ray images and classify them into several disease types or severities in radiology.3,4 Zhou et al recently demonstrated that a DL model based on ultrasound images could provide an early diagnostic strategy for lymph node metastasis in patients with breast cancer.5

The developments in the medical imaging field alone suggest we will see significant changes in clinical practice, as Rajpurkar et al stated. “We hope that this technology can improve healthcare delivery and increase access to medical imaging expertise in parts of the world where access to skilled radiologists is limited,” they wrote.

Risk Stratification in the Treatment of GISTs

Kang et al recently published the results of their first-ever study that investigates whether DL could be used as a tool to predict risk stratification in gastrointestinal stromal tumors (GISTs).6

GISTs are rare mesenchymal tumors that can originate from any part of the gastrointestinal tract. While most GISTs are benign, a small number can be quite aggressive. GISTs account for 1%-2% of all gastrointestinal tumors.7,8 

The evaluation of malignancy risk for GISTs is based on tumor size, location, and mitotic count as assessed through postoperative specimens. These factors are used by the National Institutes of Health (NIH) risk category criteria, which stratify GISTs into 4 risk categories: very low, low, intermediate, and high-risk tumors.

Preoperative risk classification can be a valuable tool for evaluating the adequacy of surgical resection and the need for additional treatment in patients with GIST. In this context, contrast-enhanced computed tomography (CT) is widely recognized as the main imaging method for diagnosing, characterizing, and evaluating the curative effect in GIST patients.9-12

Radiomics is a technique that uses data-characterization algorithms to extract a vast number of characteristics from medical images. These features, known as radiomic features, can reveal tumoral patterns and characteristics that the naked eye cannot detect.13 Several studies have shown that radiomics based on CT scans can be used to predict malignancy in GISTs. However, the radiomics approach is heavily reliant on handcrafted feature engineering.14-19

Therefore, Kang and colleagues sought to develop and validate a DL-based model capable of objectively predicting the risk stratification of GISTs using a large patient data set from 2 separate institutions.6

A Better Approach to Categorizing Risk?

A total of 733 patients (352 men; mean age, 59.8±10.1 years) with GISTs from January 2011 to June 2020 were enrolled in this retrospective study. Characteristics such as age, gender, tumor location, tumor size, and mitotic count, were derived from medical records.6

As a verification of the model, the malignant potential of GISTs was stratified by using the modified NIH criteria. Patients were divided into the low-malignant (very low and low risk), intermediate-malignant (intermediate risk), and high-malignant (high risk) potential groups according to risk categories. The patients underwent an abdominal contrast-enhanced CT examination, which revealed tumor sizes ranging from 10-240 mm.6 

The DL model was trained in 2 steps: 1) tumor feature extraction and tumor classification, and 2) multisequence-based feature fusion and patient diagnosis. Residual neural network (ResNet) was used to train the image data and build the neural network model. The study also involved the use of radiomics models for diagnosis, in addition to DL. This was done as a comparison to evaluate the efficacy of each method.6

The patients were split into 3 independent cohorts: the training, testing, and external validation cohorts. “In the training cohort, 141 (58.5%) were low-malignant GISTs, 43 (17.8%) were intermediate-malignant GISTs, and 57 (23.7%) were high-malignant GISTs. In the testing cohort, 61 (58.7%) were low-malignant GISTs, 18 (17.3%) were intermediate malignant GISTs, and 25 (24.0%) were high-malignant GISTs. In the external validation cohort, 137 (35.3%) were low malignant GISTs, 67 (17.3%) were intermediate-malignant GISTs, and 184 (47.4%) were high-malignant GISTs,“ the researchers reported.6

The study’s findings suggested that the DL model could accurately predict the risk classification of GISTs in the testing cohort with areas under the receiver operating characteristic curves (AUROCs)=0.90. The external validation cohort’s performance was lower, but still positive (AUROCs=0.81). The proposed DL model outperformed the radiomics model in both the testing and external validation cohorts, indicating that the DL model could mine more image features useful for assessing the risk classification in patients with GISTs.6

“Our work represents an improved approach to the assessment of risk stratification based on the CT images from patients GISTs obtained before surgery and significantly improves on current prediction methods that rely on postoperative specimens,” Kang et al wrote.

They also address the limitations of their study and indicate that the generalizability of their assessment tool (DL model) needs to be further evaluated. 

Kang et al’s DL model has the potential to serve as an important decision support tool in clinical applications for predicting risk stratification in patients with GISTs. It is just one example of physicians and data scientists working together today to create technological advances that improve patient care. The use of deep learning in medical settings has shown promising results in several therapeutic areas and may provide more accurate information about patients’ conditions in the near future.


  1. Murphy KP. Machine Learning: A Probabilistic Perspective. Cambridge, MA: The MIT Press; 2022.
  2. Deng L. Deep learning: methods and applications. Foundations and Trends in Signal Processing. 2014;7(3-4):197-387. doi:10.1561/2000000039
  3. Rajpurkar P, Irvin J, Zhu K, et al. CheXNet: radiologist-level pneumonia detection on chest x-rays with deep learning. Computing Research Repository. 2017;abs/1711.05225. doi:10.48550/arXiv.1711.05225
  4. Qin C, Yao D, Shi Y, Song Z. Computer-aided detection in chest radiography based on artificial intelligence: a survey. Biomed Eng Online. 2018;17(1):113. doi:10.1186/s12938-018-0544-y
  5. Zhou LQ, Wu XL, Huang SY, et al. Lymph node metastasis prediction from primary breast cancer US images using deep learning. Radiology. 2020;294(1):19-28. doi:10.1148/radiol.2019190372
  6. Kang B, Yuan X, Wang H, et al. Preoperative CT-based deep learning model for predicting risk stratification in patients with gastrointestinal stromal tumors. Front Oncol. 2021;11:750875. doi:10.3389/fonc.2021.750875
  7. Joensuu H, Hohenberger P, Corless CL. Gastrointestinal stromal tumour. Lancet. 2013;382(9896):973-983. doi:10.1016/S0140-6736(13)60106-3
  8. Parab TM, DeRogatis MJ, Boaz AM, et al. Gastrointestinal stromal tumors: a comprehensive review. J Gastrointest Oncol. 2019;10(1):144-154. doi:10.21037/jgo.2018.08.20
  9. Li J, Gong JF, Wu AW, Shen L. Post-operative imatinib in patients with intermediate or high risk gastrointestinal stromal tumor [published correction appears in Eur J Surg Oncol. 2011 Oct;37(10):919]. Eur J Surg Oncol. 2011;37(4):319-324. doi:10.1016/j.ejso.2011.01.005
  10. Lin JX, Chen QF, Zheng CH, et al. Is 3-years duration of adjuvant imatinib mesylate treatment sufficient for patients with high-risk gastrointestinal stromal tumor? A study based on long-term follow-up. J Cancer Res Clin Oncol. 2017;143(4):727-734. doi:10.1007/s00432-016-2334-x
  11. Inoue A, Ota S, Sato S, et al. Comparison of characteristic computed tomographic findings of gastrointestinal and non-gastrointestinal stromal tumors in the small intestine. Abdom Radiol (NY). 2019;44(4):1237-1245. doi:10.1007/s00261-018-1865-9
  12. Vernuccio F, Taibbi A, Picone D, et al. Imaging of gastrointestinal stromal tumors: from diagnosis to evaluation of therapeutic response. Anticancer Res. 2016;36(6):2639-2648.
  13. Yip SSF, Liu Y, Parmar C, et al. Associations between radiologist-defined semantic and automatically computed radiomic features in non-small cell lung cancer. Sci Rep. 2017;7(1):3519. doi:10.1038/s41598-017-02425-5
  14. Wang C, Li H, Jiaerken Y, et al. Building CT radiomics-based models for preoperatively predicting malignant potential and mitotic count of gastrointestinal stromal tumors. Transl Oncol. 2019;12(9):1229-1236. doi:10.1016/j.tranon.2019.06.005
  15. Zhang L, Kang L, Li G, et al. Computed tomography-based radiomics model for discriminating the risk stratification of gastrointestinal stromal tumors. Radiol Med. 2020;125(5):465-473. doi:10.1007/s11547-020-01138-6
  16. Chen T, Ning Z, Xu L, et al. Radiomics nomogram for predicting the malignant potential of gastrointestinal stromal tumours preoperatively. Eur Radiol. 2019;29(3):1074-1082. doi:10.1007/s00330-018-5629-2
  17. Feng C, Lu F, Shen Y, et al. Tumor heterogeneity in gastrointestinal stromal tumors of the small bowel: volumetric CT texture analysis as a potential biomarker for risk stratification. Cancer Imaging. 2018;18(1):46. doi:10.1186/s40644-018-0182-4
  18. Chen Z, Xu L, Zhang C, et al. CT radiomics model for discriminating the risk stratification of gastrointestinal stromal tumors: a multi-class classification and multi-center study. Front Oncol. 2021;11:654114. doi:10.3389/fonc.2021.654114
  19. Berenguer R, Pastor-Juan MDR, Canales-Vázquez J, et al. Radiomics of CT features may be nonreproducible and redundant: influence of CT acquisition parameters. Radiology. 2018;288(2):407-415. doi:10.1148/radiol.2018172361