CT Radiomics for Tissue Prediction

Exploratory analysis of ex vivo liver specimen

21

Biopsy Samples

At different CT locations

12

CT Features

Intensity-based radiomics

3

Models

Tumor, necrosis, fibrosis

0.94

Avg R²

Correlation observed

Overview

Study goals and context

Workflow

Methods and process

Results

Model performance

Discussion

Progress and notes

Project Overview

Study Goal

Explore if CT radiomic features correlate with tissue composition
Test feasibility of predicting tumor, necrosis, and fibrosis from imaging
Validate workflow for imaging-pathology correlation

Resected left hepatic lobe

CT imaging of specimen

H&E pathology annotation

Approach

21 biopsy samples collected at different locations across ex vivo liver specimen
H&E staining with pathologist annotations for ground truth (tumor, necrosis, fibrosis)
12 CT imaging features extracted from voxels at each biopsy location (intensity-based radiomics)
Multiple linear regression models built to predict tissue composition from CT features

Annotated Biopsy Specimen Histology

H&E stained liver biopsy specimens with pathologist annotations (yellow outlines) showing tissue regions used for ground truth classification.

Specimen Annotation Comparison

Side-by-side comparison of H&E stained specimen and pathologist annotations showing tumor regions (red) and necrosis/hemorrhage (green).

Note: Exploratory study with n=21 biopsy samples. Demonstrates methodology, not validated clinical findings.

Analysis Workflow

Click each step for details

1

Specimen & CT Imaging

▼

Resected left hepatic lobe from NCI Lab
In vivo and ex vivo CT scans acquired
Segmentation and 3D reconstruction

2

Biopsy Collection (21 samples)

▼

21 core needle biopsy samples collected at different locations
Distributed across tumor and surrounding tissue
CT scanned with needles in place for precise 3D coordinate registration

3

Pathology Analysis (Ground Truth)

▼

H&E staining on all 21 biopsy samples
Pathologist annotated tumor, necrosis, fibrosis
Calculated % composition for each tissue type at each location

4

CT Feature Extraction (49 features)

▼

Defined 16×16×3 voxel ROI at each of the 21 biopsy locations
49 intensity-based CT radiomic features extracted per location
Features computed per IBSI guidelines: intensity statistics, histogram metrics, texture descriptors

5

Feature Reduction (49 → 12)

▼

Pair-wise correlation filtering applied to avoid multicollinearity
Reduced from 49 to 12 independent CT imaging features
Final matrix: 21 biopsy samples × 12 CT radiomic features

6

Model Building

▼

Built 3 multiple linear regression models
Model 1: 12 CT features → Predict % Tumor
Model 2: 12 CT features → Predict % Necrosis
Model 3: 12 CT features → Predict % Fibrosis

Results

R²

0.95

Adjusted R²

0.90

RMSE

5.16%

p-value

1.1×10^-4

R²

0.92

Adjusted R²

0.84

RMSE

5.92%

p-value

3.2×10^-4

R²

0.94

Adjusted R²

0.90

RMSE

1.67%

p-value

3.5×10^-6

Key Observations

All three models show strong correlations (R² 0.92-0.95) in this dataset
12 CT intensity-based features captured tissue composition signal
Proof-of-concept for CT-to-pathology correlation workflow
Results are exploratory with 21 biopsy samples, not clinically validated

Discussion

Progress So Far

Complete workflow from biopsy collection to modeling validated
12 CT radiomic features correlate with pathology at 21 biopsy locations
Methodology is reusable for other imaging-pathology datasets
Dataset available for analysis: 27 patients with CT-guided liver biopsies, of which 17 patients have tumor specimen

Notes

Single specimen: Cannot assess generalization across patients
Ex vivo to in vivo: Will our findings on ex vivo translate well onto in vivo imaging?

Discussion Questions

How to make use of the remaining funds?
Are there datasets that could benefit from the CT radiomics approach?
This framework is also being applied to the ultrasound-guided biopsy data registered on MRI (ultrasound-MRI biopsy dataset available on TCIA)