Sensitivity of an AI method for [18F]FDG PET/CT outcome prediction of diffuse large B-cell lymphoma patients to image reconstruction protocols

Abstract

Background: Convolutional neural networks (CNNs), applied to baseline [¹⁸F]-FDG PET/CT maximum intensity projections (MIPs), show potential for treatment outcome prediction in diffuse large B-cell lymphoma (DLBCL). The aim of this study is to investigate the robustness of CNN predictions to different image reconstruction protocols. Baseline [¹⁸F]FDG PET/CT scans were collected from 20 DLBCL patients. EARL1, EARL2 and high-resolution (HR) protocols were applied per scan, generating three images with different image qualities. Image-based transformation was applied by blurring EARL2 and HR images to generate EARL1 compliant images using a Gaussian filter of 5 and 7 mm, respectively. MIPs were generated for each of the reconstructions, before and after image transformation. An in-house developed CNN predicted the probability of tumor progression within 2 years for each MIP. The difference in probabilities per patient was then calculated between both EARL2 and HR with respect to EARL1 (delta probabilities or ΔP). We compared these to the probabilities obtained after aligning the data with ComBat using the difference in median and interquartile range (IQR).

Results: CNN probabilities were found to be sensitive to different reconstruction protocols (EARL2 ΔP: median = 0.09, interquartile range (IQR) = [0.06, 0.10] and HR ΔP: median = 0.1, IQR = [0.08, 0.16]). Moreover, higher resolution images (EARL2 and HR) led to higher probability values. After image-based and ComBat transformation, an improved agreement of CNN probabilities among reconstructions was found for all patients. This agreement was slightly better after image-based transformation (transformed EARL2 ΔP: median = 0.022, IQR = [0.01, 0.02] and transformed HR ΔP: median = 0.029, IQR = [0.01, 0.03]).

Conclusion: Our CNN-based outcome predictions are affected by the applied reconstruction protocols, yet in a predictable manner. Image-based harmonization is a suitable approach to harmonize CNN predictions across image reconstruction protocols.

Keywords: Convolutional neural networks; Diffuse large B-cell lymphoma; PET; Reconstruction.

Fig. 1

Workflow overview. A Generation of original probabilities from whole-body PET scans. The MIPs are generated from the PET scan through the preprocessing tool. The CNN is then used to predict 2-year TTP probabilities. This is done for each of the 3 reconstructed images for all patients. B Generation of image-transformed probabilities from filtered whole-body PET scans. A Gaussian filter is applied to the EARL2 and HR scans to obtain images that resemble EARL1-compliant images. The preprocessing tool is used to generate the MIPs from the transformed scans and the CNN is then used to predict the corresponding 2-year TTP probabilities. C Generation of ComBat probabilities from whole-body PET scans. To obtain the ComBat-transformed probabilities, ComBat is applied to the generated original probabilities

Fig. 2

MIP images for the same patient for the three reconstruction protocols. (A) MIP images with their corresponding CNN predictions (P). From left to right: EARL1, EARL2 and HR. (B) MIP images after image-based transformation (except for EARL1). From left to right: EARL1, EARL2 and HR. Predictions from the original MIPs are shown in red, from the transformed or ‘blurred’ MIPs in green and after ComBat transformation in blue

Fig. 3

Regression lines for CNN probabilities with EARL1 as reference. A EARL2 original probabilities (red), EARL2 probabilities after image-based transformation (green) and EARL2 probabilities after ComBat transformation (blue) compared to EARL1. B HR original probabilities (red), HR probabilities after image-based transformation (green) and HR probabilities after ComBat transformation (blue) compared to EARL1. The probability values are closer to the line of identity (gray-dashed line) for both EARL2 and HR values after both transformations

Fig. 4

Bland–Altman plots. A Bland–Altman plot for EARL2 and EARL1 probabilities, before transformation (in red) and after both image (in green) and ComBat transformation (in blue). B Bland–Altman plot for HR and EARL1 probabilities, before transformation (in red) and after both image (in green) and ComBat transformation (in blue)