Why Predicting Survival in Pancreatic Cancer Is So Difficult

PLoS One 2017 AI 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Why Predicting Survival in Pancreatic Cancer Is So Difficult

Pancreatic cancer is one of the most lethal cancers, with a five-year survival rate below 10%. One major challenge is that outcomes vary considerably between patients even after surgery, making it hard to identify who will fare better or worse without waiting years to find out.

Physicians currently rely on clinical variables like tumor stage and size to estimate prognosis, but these measures often miss the complexity inside the tumor itself. Two tumors that look the same on a scan may behave very differently at the cellular level.

This study explored whether CT image texture analysis -- a technique that quantifies subtle patterns in grayscale intensity across an image -- could provide additional predictive information beyond what clinicians can visually assess, helping to identify patients unlikely to survive two years after treatment.

TL;DR: Standard clinical staging misses important survival-relevant differences between pancreatic tumors, motivating the search for objective imaging-based predictors.
Pages 2-4
Extracting Texture Features from Pre-Treatment CT Scans

The researchers worked with 35 pancreatic cancer patients enrolled in a phase II clinical trial at Memorial Sloan Kettering Cancer Center. Pre-treatment CT scans were collected, and a three-dimensional region of interest was manually drawn around each tumor by an expert radiologist.

From these regions, the team extracted 255 texture features using six mathematical frameworks: Gray Level Co-occurrence Matrix (GLCM), Run Length Matrix (RLM), Local Binary Patterns (LBP), Fractal Dimension (FD), Intensity Histogram (IH), and Autocorrelation Matrix (ACM). Each method captures a different aspect of how pixel intensities relate to one another spatially.

The fMRMR (fuzzy minimum-redundancy maximum-relevance) algorithm was applied to select the most informative and least redundant features from this large set. Selected features were then used to train a naive Bayes classifier to predict whether a patient would survive two years after treatment.

Because the dataset was small, the team used leave-one-out cross-validation, where each patient is held out in turn as a test case while the model trains on the remaining patients. This approach maximizes use of the available data while providing an honest performance estimate.

TL;DR: 255 texture features from CT tumor regions were extracted using six methods, then filtered and used to train a classifier predicting two-year survival in 35 patients.
Pages 5-8
ACM Features Outperform Other Texture Methods

Across all six texture feature families, the Autocorrelation Matrix (ACM) features consistently provided the best predictive performance. The top ACM feature subset (ACM2) achieved an area under the ROC curve (AUC) of 0.90 and a classification accuracy of 82.86%.

By comparison, standard clinical parameters alone provided much weaker discrimination. This gap suggests the imaging texture features are capturing meaningful biological variation in tumor composition that is invisible to the naked eye and not encoded in staging information.

The model identified patterns related to tumor heterogeneity -- the degree to which different parts of the tumor have different cell densities, vascularity, or necrosis. Tumors with higher heterogeneity tended to be associated with shorter survival, consistent with the idea that chaotic tumor architecture reflects more aggressive biology.

Results were reproducible across different subsets of features and feature selection parameters, lending confidence that the findings were not driven by a single feature or artifact of a particular analysis configuration.

TL;DR: The ACM texture feature family achieved an AUC of 0.90 for predicting two-year survival, significantly outperforming clinical variables alone.
Pages 9-11
What Texture Features Reveal About Tumor Biology

Texture analysis works on the assumption that the spatial arrangement of pixel intensities in a CT image reflects underlying tissue structure. Tumors with irregular vascularity, necrotic cores, or fibrotic stroma produce distinctive patterns that algorithms can detect even when those patterns are not obvious to a radiologist.

The strong performance of ACM features is notable because ACM captures second-order statistical relationships between pixel pairs across multiple distances and orientations. This multi-scale sensitivity may help it detect heterogeneity at both fine and coarse spatial levels within the tumor.

The study supports a growing body of evidence that radiomics -- the systematic extraction of quantitative features from medical images -- can serve as a non-invasive window into tumor biology. Features that predict survival may ultimately reflect gene expression patterns or proteomic states that drive aggressive behavior.

TL;DR: ACM texture features likely detect multi-scale tumor heterogeneity in CT images, acting as a non-invasive surrogate for aggressive tumor biology.
Pages 11-12
Study Limitations and the Challenge of Small Cohorts

The most significant limitation of this study is the small sample size of 35 patients. With so few cases, even well-designed cross-validation procedures can produce optimistic performance estimates that fail to hold up in larger, independent datasets.

All patients came from a single institution and a single clinical trial, which may not represent the diversity of pancreatic cancer patients seen in general clinical practice. Differences in CT scanner hardware, reconstruction protocols, and tumor segmentation practices between institutions could all affect texture feature values.

The study treated two-year survival as a binary outcome, collapsing a continuous survival time into a threshold. While practical, this approach discards granular prognostic information and makes the results harder to compare with studies using continuous survival endpoints like overall survival or progression-free survival.

TL;DR: The 35-patient single-institution cohort limits generalizability, and multicenter validation with larger samples is needed before clinical use.
Pages 12-13
Toward Imaging Biomarkers for Pancreatic Cancer Prognosis

This preliminary study demonstrates that CT texture analysis, particularly using autocorrelation matrix features, can predict two-year survival in pancreatic cancer patients with accuracy substantially better than clinical staging alone.

If validated in larger cohorts, these features could serve as imaging biomarkers that are available the moment a pre-treatment CT scan is acquired -- no additional tests, biopsies, or waiting. This would allow oncologists to better inform patients about prognosis and to stratify patients in clinical trials.

Future work should focus on prospective validation across multiple institutions, integration with other biomarkers such as genomic or serum markers, and exploration of whether texture features can also predict response to specific therapies rather than just overall survival.

TL;DR: CT texture analysis could become a non-invasive imaging biomarker for pancreatic cancer prognosis, pending validation in larger multicenter studies.
Citation: Open Access, 2017. Available at: PMC5720792.