top of page
  • Dr Rimjhim Agrawal and Varsha Raghuram

Machine Learning in Neurosurgery

The current process of decision-making in neurosurgery is based on clinical practice guidelines which cite clinical trials and case series publications. Statistical methods are utilized in most of these publications, including but not limited to: t-tests, cox-hazard model and chi-square test. They calculate the correlation between variables, but they require an assumption of determining the level of significance. Moreover, correlation does not imply causation. Statistical tests are limited in many ways; they cannot analyze non-linear variables and struggle to find correlations in larger data sets.

In the past few decades, machine learning is being explored as a new method of computational data analysis to overcome the limits of statistical tests. Machine learning uses algorithms and statistical models to analyze and make inferences from patterns in data, and can learn and adapt without explicit instruction. It is highly capable of performing exhaustive analyses even with massive amounts of non-linear data, removing two of statistics’ greatest hurdles.

There are two kinds of machine learning: supervised and unsupervised. Supervised learning aims to make the machine predict the output, and works via a training period where all possible inputs and outputs are registered. The algorithm is repeatedly tested, observing the quality of predictions. If a large percentage of its predictions are accurate to the known conclusion, it can then be used to make predictions where the outputs are not known.

A good example is a program that identifies breeds of dogs. The program is trained with a large number of pictures of dogs, and is told what the correct answer for each breed is. It “learns” the features of each type so it can accurately identify any dog picture it’s given.

In unsupervised learning, the aim is to find patterns in data rather than to predict an outcome. Pattern analysis is especially helpful in revealing pathophysiological mechanisms of disease and comprehending the way variables overlap. For example, in a glioma patient, finding the relationship between imaging characteristics, molecular biomarkers, performance score and amount of cerebral edema would be more important than predicting an outcome.

Many different mathematical algorithms can be defined for machine learning:

  1. Logistic regression is one where inputs and outputs come additively and linearly, but it does not allow for further addition of alternative variables

  2. The decision tree algorithm performs predictions from different variables even if they do not occur together, suggesting possible correlations between symptoms and underlying diseases

  3. Neural networks provide a massive amount of flexibility, allowing for their input features to be changed at will. Their input types can even be added to during data collection

  4. Bayesian networks predict the conditional dependency of random variables

  5. Linear discriminant analysis aims to find a linear combination of features that characterizes (or separates) two or more occurrences

Drage et al. (2022) discuss the effectiveness of using machine learning algorithms in predicting outcomes. The authors use colorectal carcinoma (CRC) as an example, for which microsatellite instability (MSI) is associated with prognosis, therapeutic response, and even an inherited cancer predisposition syndrome. At first, pathologists manually evaluated known characteristic morphologic features to identify patients for testing, but the data showed a lack of sensitivity to this approach. Practice guidelines shifted to universal testing, but these techniques require additional resources that are not available everywhere.

Testing shows that computers can be trained to identify MSI tumors across multiple types of diseases with different outcomes, based on just an H&E image. The sensitivity and specificity of the Convoluted Neural Network (CNN) in identifying these comes close to that of routine ancillary testing (using IHC and PCR imaging). The CNN approach proves capable of predicting gene-expression based consensus molecular subtype classification of CRC based on the H&E images, providing extra measures of intratumoral spatial heterogeneity and even proposing a new class for a subset of tumors that were unclassifiable through RNA expression profiling.

Machine learning shows potential in predicting other clinically relevant biomarkers like PD-L1 expression and relevant molecular characteristics in non-small-cell lung cancer. The authors argue that a distributable algorithm with good performance could become a cost-effective screening tool.

Image credit: Michael Drage,

Dundar et al. (2022) worked on “individual-safe, intracranial approaches by introducing functional anatomical structures and pathological areas to artificial intelligence”. They used preoperative magnetic resonance images of patients with deeply located brain tumors for planning. The voxel values of these selected regions in cranial MR sequences were extracted and labeled; tumor tissue was segmented as the target.

A Q-learning algorithm was run on these voxel values, on optimal paths extracted by a new heuristic-based path-planning algorithm. The Q-learning algorithm was run to list the cortico-tumoral pathways that can ensure maximum removal of tumor tissue with the least effect on functional anatomical tissue. The results were positive: the most suitable intracranial areas for incision and surgery were located, and the cortico-tumoral pathways were revealed from these optimal points using the Q-algorithm.

Image credit:

A 2019 survey by the European Association of Neurosurgical Societies (EANS) and the Congress of Neurosurgeons (CNS) explored how many neurosurgeons use or advocate for ML in surgical planning. A short online questionnaire was sent to 7280 neurosurgeons across the globe, 362 responses were received, largely from surgeons in Europe and North America. Of these, 103 surgeons reported using ML in their clinical practice (28.5%), and 31.1% used it in research. The most common application of machine learning was to predict outcomes and complications and interpret imaging.

Machine learning could set a precedent for powerful, accurate and fast interpretation of existing data to make inferences and reach predictions that would otherwise require manual work. It shows a degree of resistance to human error and bias, and can adapt to new learnings with experience. Machine learning promises to become more accurate in its decision-making, improve clinical and surgical management as well as pre-surgical planning on a large scale.

Have we piqued your interest in machine learning? You can read our white papers here and publications here.


Celtikci, E. (2017, May 2). A systematic review on machine learning in Neurosurgery: The future of decision-making in patient care. Turkish neurosurgery. Retrieved October 10, 2022, from

Staartjes, V. E., Stumpo, V., Kernbach, J. M., Klukowska, A. M., Gadjradj, P. S., Schröder, M. L., Veeravagu, A., Stienen, M. N., van Niftrik, C. H. B., Serra, C., & Regli, L. (2020, August 18). Machine learning in neurosurgery: A global survey - acta neurochirurgica. SpringerLink. Retrieved October 10, 2022, from

Drage, M., Khosla, A., Walk, E., Mountain, V., & Lin, M. (n.d.). Machine learning in pathology: The potential to predict the future... College of American Pathologists. Retrieved October 10, 2022, from

Dundar, T. T., Yurtsever, I., Pehlivanoglu, M. K., Yildiz, U., Eker, A., Demir, M. A., Mutluer, A. S., Tektaş, R., Kazan, M. S., Kitis, S., Gokoglu, A., Dogan, I., & Duru, N. (2022). Machine Learning-Based Surgical Planning for Neurosurgery: Artificial Intelligent Approaches to the Cranium. Frontiers in surgery, 9, 863633.

Barreto, S. (2022, November 5). Real-life examples of supervised learning and unsupervised learning. Baeldung on Computer Science. Retrieved December 21, 2022, from

bottom of page