A95-10 Deep Learning for Digital Histopathological Classification of Fibrotic Lung Disease: A COLDICE Substudy

Abstract Rationale Interstitial lung diseases (ILD) are heterogeneous conditions, with biopsy occasionally indicated to inform multidisciplinary discussion (MDD) for diagnosis. Variation in histopathological interpretation, particularly for small tissue samples, may impact MDD classification and diagnostic confidence. A deep-learning algorithm was developed and evaluated using biopsies obtained from the COLDICE (Cryobiopsy versus Open Lung Biopsy in the Diagnosis of Interstitial Lung Disease Alliance) study as a support tool for histopathologic interpretation. Methods Hematoxylin and eosin-stained slides from 130 samples (65 paired surgical lung biopsies SLB and transbronchial lung cryobiopsies TBLC) from COLDICE participants (PMID31578168) were digitised at 40x magnification. Whole slide image (WSI) tiles were separated into independent training, validation and holdout set cohorts per biopsy type. A novel semi-supervised deep-learning model was developed as follows: 1. Representative normal and abnormal tiles were extracted from WSIs using the VIPR (Validated Identification of Pre-qualified Regions) algorithm to enrich for both classes; 2. Tile cohorts were adjudicated by an expert pathologist and classified as: a) definite/probable usual interstitial pneumonia (UIP), b) indeterminate for UIP, c) alternative diagnostic features, or d) no significant abnormality; 3. A standard Python tile-augmentation library was employed to generate variants of allocated training tiles; and 4) Augmented tiles trained a convolutional neural network (CNN) to assign computer-generated UIP likelihood scores (0-1). Further CNN model refinement was undertaken using a ConvNeXt-Tiny backbone. Algorithm prediction accuracy for identifying definite/probable UIP, as per consensus histopathological COLDICE biopsy classification, was assessed along with sensitivity, specificity, and area under the receiver operating curve (AUROC), at a 0.5 threshold in holdout set cases. Results Definite/probable UIP was determined by consensus in 41/65 (63%) TBLC and 39/65 (60%) SLB specimens. Tiles generated from biopsy slides were allocated to cohorts as follows: SLB: training set n = 361, validation set n = 107, holdout set n = 107; TBLC: training set n = 175, validation set n = 34, holdout set n = 50. Algorithm-generated mean UIP likelihood score was 53.9 (SD19.0) for TBLC and 68.0 (SD13.2) for SLB, with 31/50 TBLC tiles and 96/108 SLB tiles positively classified. Pooled sensitivity and specificity of the algorithm for predicting UIP in SLB samples were 97.8% and 67.7%, respectively; accuracy 93.5%; AUROC 0.986. TBLC sensitivity was 83.3%, specificity 92.9%, accuracy 86.0%, and AUROC 0.972 (figure 1). Conclusions A semi-supervised deep learning model provided highly accurate and reproducible UIP classification using either SLB or TBLC samples. Following further validation, this tool has the potential to support diagnostic decision-making in less experienced centers. This abstract is funded by: The COLDICE study was funded by Lung Foundation Australia, University of Sydney, Hunter Medical Research Institute, Erbe Elektromedizin, Medtronic, Cook Medical, Rymed, Karl-Storz, Zeiss, and Olympus