No prognostic tool, no matter how refined, does everything.
For instance, molecular assays that gauge cancer-related signatures are challenged by their inability to factor in tissue architecture and the results are confounded by genomic information from the different types of cells inside the tumor other than cancer cells. Meanwhile, traditional histopathological assessments are good at gauging tissue architecture and differentiating cellular heterogeneity, but mostly provide qualitative tumor data and are too time consuming to be applied in large-scale studies.
Recognizing these weaknesses, researchers led by Yinyin Yuan of Cancer Research UK decided to combine histopathological and gene expression analysis to show that quantitative image analysis of the cellular environment inside tumors can bolster the ability of genomic profiling to predict survival in breast cancer patients. "All technologies have some sort of weakness. That's why when we combined two types of assays — image and microarray — we get a more reliable readout," Yuan says.
As they report in Science Translational Medicine, Yuan and her colleagues gathered histopathological information from hematoxylin and eosin-stained images as well as gene expression and copy-number variation data on a discovery set of 323 samples and on a validation set of 241 samples from patients with estrogen receptor-negative breast cancer. Using the discovery sample set, the investigators developed an image-processing method to differentiate the cells inside tumor samples as cancerous, lymphocytic, or stromal. They then tested this technique on the validation sample.
Once Yuan and colleagues had an accurate picture of the types of cells in the tumor samples, they used image analysis to correct copy-number data — as it is influenced by cellular heterogeneity — and developed an algorithm to determine patients' HER2 status better than copy-number analysis can.
Using the image-processing method, the researchers stratified the discovery and validation sample sets into lymphocytic infiltration-high and lymphocytic infiltration-low groups — as past studies have suggested that high lymphocytic infiltration is linked to better patient outcomes.
When the image analysis was compared to the pathological scores of the samples, the discovery set showed no difference in patient outcomes, but the assessments disagreed with regard to the outcomes of the lymphocytic infiltration-low group in the validation cohort.
Hypothesizing that integrating the gene expression signatures and quantitative image analysis would improve survival prediction, the study investigators combined them. "The gene expression classifier had 67 percent cross-validation accuracy in predicting disease-specific deaths, the image-based classifier had 75 percent, and the integrated classifier reached 86 percent," the study authors write.
Finally, Yuan and her colleagues applied the image analysis to develop a quantitative score that determines whether specific types of cells are tightly clustered — a high score — or are randomly scattered — a low score. In stromal cells, this approach could discern that breast cancer patients with a high or low score had a "significantly better outcome" than patients whose scores fell in the medium range.
Ultimately, Yuan and her colleagues show that their image processing avoids the biases of manual pathological assessments and accurately quantifies cellular composition and tissue architecture not accounted for by molecular tests. The researchers' computational approach is also faster than traditional pathological techniques. "These two sets of samples can be done in a day," Yuan says.
According to the study authors, the limitation of the image processing technique is, of course, that it requires matched molecular and image data.