RETURN TO ABSTRACT LISTING


Objective:
The International Staging System (ISS) for multiple myeloma categorizes patients into three stages, with a median survival of 62 months, 45 months, and 29 months for stage I, II, and II, respectively. The aim of this study is to develop a deep learning model that predicts the ISS stage of patients with new diagnoses of multiple myeloma (NDMM) from low-dose whole-body CT skeletal surveys (LDSS).

Materials and Methods:
LDSS from patients with NDMM were included in the study. The dataset comprised 43 ISS-I, 66 ISS-II, and 75 ISS-III subjects. The axial soft tissue kernel series (2-mm slice thickness) were resized to 256 x 256 x 800 voxels and used for model development. To ensure a balanced distribution of labels, we divided our training dataset into five folds at the subject level. To evaluate the reliability of our model, we conducted five-fold cross-validation. Using a 3DDensenet121, we trained the model to classify the three ISS stages. We used the weighted cross-entropy loss to compensate for any imbalance between the classes, with AdamW as the optimizer and CosineAnealing as the learning rate scheduler. Algorithm performance was evaluated based on the area under the receiver operating characteristic curve (AUC ROC) for each validation fold.

Results:
The data we collected comprised 184 patients, with a median age of 68.85 (with an interquartile range [IQR] of 10.51) and a women-to-men ratio of 1:1.9. The classification performance of our model on determining the ISS of NDMM from LDSS was an AUC of 0.82, with a SD of 0.05.

Conclusion:
We developed a deep-learning model that can predict the ISS of NDMM from LDSS with good accuracy. Further model refinement and external validation are necessary to determine the generalizability of this algorithm. LDSS is the recommended initial imaging test for patients with NDMM to determine the presence and extent of bone disease. An accurate assessment of the ISS stage adds to the value of this diagnostic test.