Wookjin Choi

Wookjin Choi, Navdeep Dahiya, Saad Nadeem

Spiculations/lobulations, sharp/curved spikes on the surface of lung nodules, are good predictors of lung cancer malignancy and hence, are routinely assessed and reported by radiologists as part of the standardized Lung-RADS clinical scoring criteria. Given the 3D geometry of the nodule and 2D slice-by-slice assessment by radiologists, manual spiculation/lobulation annotation is a tedious task and thus no public datasets exist to date for probing the importance of these clinically-reported features in the SOTA malignancy prediction algorithms. As part of this paper, we release a large-scale Clinically-Interpretable Radiomics Dataset, CIRDataset, containing 956 radiologist QA/QC'ed spiculation/lobulation annotations on segmented lung nodules from two public datasets, LIDC-IDRI (N=883) and LUNGx (N=73). We also present an end-to-end deep learning model based on multi-class Voxel2Mesh extension to segment nodules (while preserving spikes), classify spikes (sharp/spiculation and curved/lobulation), and perform malignancy prediction. Previous methods have performed malignancy prediction for LIDC and LUNGx datasets but without robust attribution to any clinically reported/actionable features (due to known hyperparameter sensitivity issues with general attribution schemes). With the release of this comprehensively-annotated CIRDataset and end-to-end deep learning baseline, we hope that malignancy prediction methods can validate their explanations, benchmark against our baseline, and provide clinically-actionable insights. Dataset, code, pretrained models, and docker containers are available at https://github.com/nadeemlab/CIR.

Zubia Naz, Farhan Asghar, Muhammad Ishfaq Hussain et al.

Automated medical image captioning translates complex radiological images into diagnostic narratives that can support reporting workflows. We present a Swin-BART encoder-decoder system with a lightweight regional attention module that amplifies diagnostically salient regions before cross-attention. Trained and evaluated on ROCO, our model achieves state-of-the-art semantic fidelity while remaining compact and interpretable. We report results as mean$\pm$std over three seeds and include $95\%$ confidence intervals. Compared with baselines, our approach improves ROUGE (proposed 0.603, ResNet-CNN 0.356, BLIP2-OPT 0.255) and BERTScore (proposed 0.807, BLIP2-OPT 0.645, ResNet-CNN 0.623), with competitive BLEU, CIDEr, and METEOR. We further provide ablations (regional attention on/off and token-count sweep), per-modality analysis (CT/MRI/X-ray), paired significance tests, and qualitative heatmaps that visualize the regions driving each description. Decoding uses beam search (beam size $=4$), length penalty $=1.1$, $no\_repeat\_ngram\_size$ $=3$, and max length $=128$. The proposed design yields accurate, clinically phrased captions and transparent regional attributions, supporting safe research use with a human in the loop.

Mumtaz Hussain Soomro, Hamidreza Nourzadeh, Victor Gabriel Leandro Alves et al.

A 3D deep learning model (OARnet) is developed and used to delineate 28 H&N OARs on CT images. OARnet utilizes a densely connected network to detect the OAR bounding-box, then delineates the OAR within the box. It reuses information from any layer to subsequent layers and uses skip connections to combine information from different dense block levels to progressively improve delineation accuracy. Training uses up to 28 expert manual delineated (MD) OARs from 165 CTs. Dice similarity coefficient (DSC) and the 95th percentile Hausdorff distance (HD95) with respect to MD is assessed for 70 other CTs. Mean, maximum, and root-mean-square dose differences with respect to MD are assessed for 56 of the 70 CTs. OARnet is compared with UaNet, AnatomyNet, and Multi-Atlas Segmentation (MAS). Wilcoxon signed-rank tests using 95% confidence intervals are used to assess significance. Wilcoxon signed ranked tests show that, compared with UaNet, OARnet improves (p<0.05) the DSC (23/28 OARs) and HD95 (17/28). OARnet outperforms both AnatomyNet and MAS for DSC (28/28) and HD95 (27/28). Compared with UaNet, OARnet improves median DSC up to 0.05 and HD95 up to 1.5mm. Compared with AnatomyNet and MAS, OARnet improves median (DSC, HD95) by up to (0.08, 2.7mm) and (0.17, 6.3mm). Dosimetrically, OARnet outperforms UaNet (Dmax 7/28; Dmean 10/28), AnatomyNet (Dmax 21/28; Dmean 24/28), and MAS (Dmax 22/28; Dmean 21/28). The DenseNet architecture is optimized using a hybrid approach that performs OAR-specific bounding box detection followed by feature recognition. Compared with other auto-delineation methods, OARnet is better than or equal to UaNet for all but one geometric (Temporal Lobe L, HD95) and one dosimetric (Eye L, mean dose) endpoint for the 28 H&N OARs, and is better than or equal to both AnatomyNet and MAS for all OARs.

Sanghoon Lee, Colton Farley, Simon Shim et al.

Detecting cancer manually in whole slide images requires significant time and effort on the laborious process. Recent advances in whole slide image analysis have stimulated the growth and development of machine learning-based approaches that improve the efficiency and effectiveness in the diagnosis of cancer diseases. In this paper, we propose an unsupervised learning approach for detecting cancer in breast invasive carcinoma (BRCA) whole slide images. The proposed method is fully automated and does not require human involvement during the unsupervised learning procedure. We demonstrate the effectiveness of the proposed approach for cancer detection in BRCA and show how the machine can choose the most appropriate clusters during the unsupervised learning procedure. Moreover, we present a prototype application that enables users to select relevant groups mapping all regions related to the groups in whole slide images.

Sadegh Riyahi, Wookjin Choi, Chia-Ju Liu et al.

Quantification of local metabolic tumor volume (MTV) chan-ges after Chemo-radiotherapy would allow accurate tumor response evaluation. Currently, local MTV changes in esophageal (soft-tissue) cancer are measured by registering follow-up PET to baseline PET using the same transformation obtained by deformable registration of follow-up CT to baseline CT. Such approach is suboptimal because PET and CT capture fundamentally different properties (metabolic vs. anatomy) of a tumor. In this work we combined PET and CT images into a single blended PET-CT image and registered follow-up blended PET-CT image to baseline blended PET-CT image. B-spline regularized diffeomorphic registration was used to characterize the large MTV shrinkage. Jacobian of the resulting transformation was computed to measure the local MTV changes. Radiomic features (intensity and texture) were then extracted from the Jacobian map to predict pathologic tumor response. Local MTV changes calculated using blended PET-CT registration achieved the highest correlation with ground truth segmentation (R=0.88) compared to PET-PET (R=0.80) and CT-CT (R=0.67) registrations. Moreover, using blended PET-CT registration, the multivariate prediction model achieved the highest accuracy with only one Jacobian co-occurrence texture feature (accuracy=82.3%). This novel framework can replace the conventional approach that applies CT-CT transformation to the PET data for longitudinal evaluation of tumor response.

Wookjin Choi, Saad Nadeem, Sadegh Riyahi et al.

Spiculations are important predictors of lung cancer malignancy, which are spikes on the surface of the pulmonary nodules. In this study, we proposed an interpretable and parameter-free technique to quantify the spiculation using area distortion metric obtained by the conformal (angle-preserving) spherical parameterization. We exploit the insight that for an angle-preserved spherical mapping of a given nodule, the corresponding negative area distortion precisely characterizes the spiculations on that nodule. We introduced novel spiculation scores based on the area distortion metric and spiculation measures. We also semi-automatically segment lung nodule (for reproducibility) as well as vessel and wall attachment to differentiate the real spiculations from lobulation and attachment. A simple pathological malignancy prediction model is also introduced. We used the publicly-available LIDC-IDRI dataset pathologists (strong-label) and radiologists (weak-label) ratings to train and test radiomics models containing this feature, and then externally validate the models. We achieved AUC$=$0.80 and 0.76, respectively, with the models trained on the 811 weakly-labeled LIDC datasets and tested on the 72 strongly-labeled LIDC and 73 LUNGx datasets; the previous best model for LUNGx had AUC$=$0.68. The number-of-spiculations feature was found to be highly correlated (Spearman's rank correlation coefficient $ρ= 0.44$) with the radiologists' spiculation score. We developed a reproducible and interpretable, parameter-free technique for quantifying spiculations on nodules. The spiculation quantification measures was then applied to the radiomics framework for pathological malignancy prediction with reproducible semi-automatic segmentation of nodule. Using our interpretable features (size, attachment, spiculation, lobulation), we were able to achieve higher performance than previous models.