Bone Marrow Cytology-based Artificial Intelligence Model for Detection and Prognosis of Neuroblastoma

1. Dataset processing. In this study, the model including training set, validation set and test sets. During the model construction, the investigators process the training in internal cohort by randomly assign the enrolled participants to the training set and validation set in a 8:2 ratio. The investigators use the enrolled participants of external cohorts as test sets to evaluate the model performance independently, and then select the best model for future use.

The Ethics Committee of Xinhua Hospital Affiliated to Shanghai Jiaotong University School of Medicine apaproved this study (XHEC-C-2024-023-2). 2. image acquisition. In the AI model, the investigators use bone marrow smears of enrolled participants for cytological evaluation and image collection. During cell classification and metastasis detection, the experienced pathologists complete the bone marrow smear analysis, metastatic NB clusters in bone marrow usually exhibited aggregated round atypical cells with high nucleus/cytoplasmic ratio. The investigators scan stained bone marrow smear at 40 × magnification for digital whole slide imaging (WSI). The investigators segment WSIs into smaller patches as 512 × 512 pixels tiles and apply the Vahadane method to normalize the color of small tiles. 3. Deep learning training. In the feature extraction of bone marrow cytological image, the process including two tiers of predictions: patch-level and WSI-level predictions.

For patch-level predictions, the investigators carry out label predictions and their respective probabilities for all patches. The investigators apply model in the deep learning process as follows: recognized neural network (CNN)-resnet50 and Vision Transformer (ViT). The parameter configurations in the model as follows: optimizer-SGD, loss function-softmax cross-entropy, with a batch size of 64.

For WSI-level predictions, the investigators use a multi-instance learning (MIL) algorithm to aggregate dispersed patch-level features to WSI-level features. During MIL for WSI fusion, the investigators perform WSI-level predictions with Patch Likelihood Histogram (PLH) pipeline and Bag of Words (BoW) pipeline in combination. Subsequently, the investigators get the WSI-level prediction as final representations of participant for subsequent analytical operations. 4. Signature building. In the procession of feature selection, the investigators use LASSO (Least Absolute Shrinkage and Selection Operator) feature screening to determine the final WSI-level features of the bone marrow cytology. These selected features were then subjected to machine learning methods to develop AI model. the investigators apply several machine learning algorithms to predict metastasis state of bone marrow and/or bone in participants, such as support vector machines (SVM), Logistic regression (LR), tree-based models, such as random forests and extremely randomized trees (ExtraTrees), extreme Gradient Boosting (XGBoost), and light gradient boosting machine (LightGBM), as well as multilayer perceptron (MLP) to develop our models.

In the prognosis model, the investigators e use Cox models to construct the survival model with bone marrow cytological signature and clinical characters. 5. Model evaluation and statistical analysis. The investigators compare the clinical characteristics of participant with independent sample t-test for continuous variables and the χ² test for discrete variables in SPSS version 22.0 (SPSS, Inc., Chicago, IL, USA). P ≤ 0.05 was considered statistically significant.