Computational textural mapping harmonises sampling variation and reveals multidimensional histopathological fingerprints.

Question

1. How can computational histopathology aid in identifying patients for immunotherapy?

2. What imaging resolution was used for digitizing TCGA and Helsinki ccRCC patient cohorts?

3. What texture classes were annotated?

4. How was texture classification achieved?

Accepted Answer

Computational histopathology can identify patients associated with poor survival or treatment sensitivity by quantifying tumour-infiltrating lymphocytes to predict immunotherapy response. This inexpensive tool can help determine the efficacy of immunotherapy as adjuvant or first-line combination therapy, improving patient outcomes. By analyzing histopathological data, computational models can identify patterns related to genomic alterations, gene expression, and viral tumorigenesis, providing valuable insights for personalized treatment strategies. The Cancer Genome Atlas (TCGA) serves as a multi-omics archive for biomedical research, offering over 10,000 diagnostic slides and associated genomic, transcriptomic, epigenomic, and clinical data from 32 cancer types. Previous studies have shown that computational histopathology can improve prognostication compared to standard staging, making it a promising tool in cancer diagnostics and prognostics.

Accepted Answer

The TCGA and Helsinki ccRCC patient cohorts were digitized mainly with an imaging resolution of ~0.25 mm/px, except for 18 samples scanned at ~0.50 mm/px. This information is crucial for researchers as it ensures consistency in the digitization process, allowing for accurate comparison and analysis of histology images across different datasets. The use of standardized imaging resolutions also facilitates the integration of data from various sources, enhancing the reliability and reproducibility of research findings.

Accepted Answer

The texture classes annotated were cancer, normal renal, stromal, red blood cells, empty background, and other textures. The classes included renal cancer (n=13,057), normal renal (n=8652), stromal (n=5460), red blood cells (n=996), empty background (n=16,026), and other textures (n=8522).

Accepted Answer

Texture classification was achieved by training a multi-class CNN using a deep residual network (ResNet) with ImageNet-pretrained ResNet-18 infrastructure. The model was adapted through transfer learning, combining the pre-trained infrastructure with a fully connected layer, ReLU activation, and a softmax layer for prediction. Training occurred at all CNN layers with the Adam optimizer, fixed learning rate, batch size, and cross-entropy loss function. The model was trained on randomly cropped 256 x 256 px tiles from annotation images, augmented with horizontal-vertical rotation. Texture masks were smoothed by sliding a 3 x 3 window with a stride of 2 over the texture map and unifying the texture class in each window. In cases of ties, cancer and stroma textures were prioritized, except if the other was cancer. The pooled texture type was randomly selected from equally occurring textures.

Accepted Answer

For two continuous variables, the Wilcoxon rank-sum test was used, while for three or more continuous variables, the Kruskal-Wallis test was applied. These non-parametric tests are suitable for comparing medians and interquartile ranges without assuming a normal distribution. The Wilcoxon rank-sum test compares the ranks of the data points between two groups, while the Kruskal-Wallis test extends this concept to multiple groups. Both tests are robust against outliers and are commonly used in biomedical research to analyze non-normally distributed data. Additionally, the Benjamini-Hochberg correction was employed to adjust p-values, reducing the likelihood of false positives in multiple comparisons. This correction method controls the false discovery rate, ensuring the reliability of the statistical analysis.

Accepted Answer

In texture-specific lymphocyte infiltration analysis, harmonisation is achieved in two steps. Firstly, the proportion of each texture area is multiplied by its relative lymphocyte proportion, reflecting lymphocyte enrichment compared to other textures. The sum of relative lymphocyte proportions in blood, cancer, normal, stroma, and other equals 100%. In the second step, the sum of normalised lymphocyte proportions is divided by the total sample area, represented as 1/4 P T t r t A t, where r is the relative lymphocyte proportion and A is the area of the texture t. To account for staining differences, samples are categorised into 'High' and 'Low' infiltration based on lymphocyte density compared to the clinical centre median density. Only centres with more than 20 samples are included in the lymphocyte analyses to mitigate the risk of nonparametric data distribution due to small batch size.

Accepted Answer

Lymphocyte infiltration is coordinated between malignant and surrounding tissue as observed in the study. The researchers used a deep learning model to identify images with high and low lymphocyte density, with the highest density unexpectedly found in the normal renal texture followed by cancer, stroma, blood, and other texture types. Intratumoural infiltration explained most of the total sample infiltration variance, and the intratumoural lymphocyte density correlated negatively with the density in the surrounding normal renal tissue and stromal texture. This indicates that lymphocyte infiltration affects all textures but unequally, possibly reflecting flow intratumoural and peritumoural areas. The results were replicated in the Helsinki dataset, confirming the generalizability of the lymphocyte classifier and the stability of the lymphocyte proportion per textures.

Accepted Answer

CTM (Canonical Texture Model) can help resolve sampling variation in TCGA histological slides by detecting up to 10% samples representing false histology or poor quality. It abrogates the association between transcriptome-based signatures and texture-specific lymphocyte proportions. By configuring algorithm training, CTM can undermine staining variance and provide a standardization step in diverse applications of computational histopathology. The approach has been demonstrated in two datasets, uncovering multifaceted connections between texture and lymphocyte density with clinical, immunological, genomic, and transcriptomic features. It can be beneficial for dataset quality control and harmonization, as well as for routine tumor grading and identifying novel biomarkers of intratumoral infiltration.

Computational textural mapping harmonises sampling variation and reveals multidimensional histopathological fingerprints.

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. How can computational histopathology aid in identifying patients for immunotherapy?

2. What imaging resolution was used for digitizing TCGA and Helsinki ccRCC patient cohorts?

3. What texture classes were annotated?

4. How was texture classification achieved?

5. What statistical tests were used for continuous variables?

6. How is lymphocyte infiltration harmonised in texture-specific analysis?

7. How is lymphocyte infiltration coordinated between malignant and surrounding tissue?

8. How can CTM help resolve sampling variation in TCGA histological slides?

Citations

Towards a general-purpose foundation model for computational pathology.

iQC: machine-learning-driven prediction of surgical procedure uncovers systematic confounds of cancer whole slide images in specific medical centers

Bioinformatics analysis of IFI6 as a novel prognostic biomarker and its correlation with immune infiltration in breast cancer

Comparison of the monoexponential and biexponential models of diffusion-weighted magnetic imaging in grading clear-cell renal cell carcinoma: a case-control study

References

Deep Residual Learning for Image Recognition

Deep Residual Learning for Image Recognition

A Survey on Transfer Learning

QuPath: Open source software for digital pathology image analysis

The Immune Landscape of Cancer

Related Papers (5)

Histopathology of nematode-induced galls of Echinochloa colonum.

The Histopathology of Maedi: A Slow, Viral Pneumonia of Sheep

Lichen striatus – a chameleon: An histopathological and immunohistological study of forty‐one cases

Distinctive histopathology of Spiroplasma eriocheiris infection in the giant river prawn Macrobrachium rosenbergii