Artificial neural network detection of pancreatic cancer from proton (1H) magnetic resonance spectroscopy patterns of plasma metabolites

Abstract

Background: Routine screening to detect silent but deadly cancers such as pancreatic ductal adenocarcinoma (PDAC) can significantly improve survival, creating an important need for a convenient screening test. High-resolution proton (1H) magnetic resonance spectroscopy (MRS) of plasma identifies circulating metabolites that can allow detection of cancers such as PDAC that have highly dysregulated metabolism.

Methods: We first acquired 1H MR spectra of human plasma samples classified as normal, benign pancreatic disease and malignant (PDAC). We next trained a system of artificial neural networks (ANNs) to process and discriminate these three classes using the full spectrum range and resolution of the acquired spectral data. We then identified and ranked spectral regions that played a salient role in the discrimination to provide interpretability of the results. We tested the accuracy of the ANN performance using blinded plasma samples.

Results: We show that our ANN approach yields, in a cross validation-based training of 170 samples, a sensitivity and a specificity of 100% for malignant versus non-malignant (normal and disease combined) discrimination. The trained ANNs achieve a sensitivity and specificity of 87.5% and 93.1% respectively (AUC: ROC = 0.931, P-R = 0.854), with 45 blinded plasma samples. Further, we show that the salient spectral regions of the ANN discrimination correspond to metabolites of known importance for their role in cancers.

Conclusions: Our results demonstrate that the ANN approach presented here can identify PDAC from 1H MR plasma spectra to provide a convenient plasma-based assay for population-level screening of PDAC. The ANN approach can be suitably expanded to detect other cancers with metabolic dysregulation.

Fig. 1. AI Analysis - neural network processing pipeline.

The schematic diagram of the processing pipeline highlights major components of the ANN (artificial neural network) approach. It primarily consists of three identical classification pathways that subdivide the three-way classification of normal, disease, and malignant (pancreatic ductal adenocarcinoma) into three separate two-way classification pathways: (I) normal vs. malignant, (II) disease vs. malignant and (III) normal vs. disease. The left half of the figure represents the training cycle and the right half represents the testing or inference cycle. The callouts (a) through (i) are included for identifying different components of the figure in the description of the ANN approach.

Fig. 2. Representative ¹H MR spectra of human plasma from normal participants (green), participants with benign pancreatic disease (blue), and participants with pancreatic ductal adenocarcinoma (red).

a Spectra were acquired using a CPMG (Carr-Purcell-Meiboom-Gill) pulse sequence (short T₂ filtering) with water pre-saturation to suppress the broad resonances from lipoproteins/albumins. b Spectra were acquired using a single pulse sequence with water pre-saturation (ZGPR). Broad peaks are from lipids, lipoproteins/albumins present in plasma. Spectral regions from 5.0 to 10.0 ppm (parts per million) were vertically magnified at 2X to better visualize the low intensity peaks in those regions. The black dotted line indicates the broad peaks that arise from macromolecules. The red dotted line identifies the flat baseline in CPMG spectra that primarily detects signal from small molecules. BCAA: branch chain amino acids; *EDTA (ethylenediaminetetraacetic acid) from blood collection tubes.

Fig. 3. Confusion-matrix results of the ANN (artificial neural network) classifications.

a–f Three-way classifications of normal (normal pancreas), disease (benign pancreatic disease), and malignant (pancreatic ductal adenocarcinoma) groups. g–l Two-way classifications with normal and disease combined versus malignant. a, b, c, g, h, i Results from the training samples. d, e, f, j, k, l Results from blinded-test samples. a, d, g, j Results from ZGPR (a single pulse sequence with water pre-saturation) spectra. b, e, h, k Results from CPMG (Carr-Purcell-Meiboom-Gill) spectra. c, f, i, l Classification accuracy comparisons between ZGPR and CPMG spectra.

Fig. 4. ANN classification performance results.

a ROC and P-R curves for the ANN (artificial neural network) malignant (pancreatic ductal adenocarcinoma) classification results. The Receiver Operating Characteristic (ROC) curves and the Precision-Recall (P-R) curves demonstrate the performance of the ANN analysis for the malignant classification in the training samples (using cross-validation) and in the blinded-test samples. The AUC (Area Under the Curves) numbers are comparable between training and blinded-test in the ROC curves and the P-R curves. A comprehensive set of multi-class ROC and P-R curves is provided in Supplementary Fig. 3. b ANN Class-probability for the blinded-test samples. The three columns of horizontal bar charts show the division of class-probability (as normal, disease and malignant) for each sample in the blinded-test set (n = 45) that were classified by the ANN analysis into normal (n = 13), disease (n = 16) and malignant (n = 16), respectively. The target (true-positive) sample sizes were 19,10,16 for normal, disease and malignant respectively. The misclassified samples were denoted by the corresponding letters as keys to indicate the correct classification. The diversity identifiers for participants other than White were included alongside their respective probability score. In the legend, Other refers to other racial groups and may include Hispanic or Latino. c, d The distribution of the class-probability in the ANN predicted samples. The distributions of class-probability percentages in the malignant versus non-malignant (normal and disease combined) classification are displayed as histogram charts for all samples that were correctly predicted as malignant (c) and, as non-malignant (d) (i.e., normal or disease) without including any of the misclassified classified samples in the respective group.

Fig. 5. Salient spectral regions that contributed to the classification as identified by the conversion of malignant samples to normal samples.

The salient spectral regions for ZGPR (a single pulse sequence with water pre-saturation) spectra (a) and CPMG (Carr-Purcell-Meiboom-Gill) spectra (b) are plotted as bar charts in the sorted order of the percent occurrences of the spectral region in the spectral substitutions that resulted in converting the classification of a malignant sample to normal. c Shift in the contributions of salient spectral regions with respect to ZGPR and CPMG spectra. The differences in the contributions of salient spectral regions in ZGPR and CPMG spectra are illustrated in this bar chart. The salient spectral regions are arranged in the x-axis in the decreasing order of ppm (parts per million) as in a standard spectral plot (i.e., 10 ppm to 0.5 ppm). The corresponding percent conversions in ZGPR and CPMG spectra (y-axis in (a, b) respectively) were compared and the differences in the percent conversion was normalized as a ratio with respect to their sum. The bars in red represent where the percent conversions in ZGPR spectra were higher than in CPMG spectra and blue bars represent percent conversions in CPMG spectra that were higher than in ZGPR spectra. The bar chart displays the shift in the contribution toward macromolecular (higher ppm) range in the case of ZGPR spectra and toward small molecule metabolites (lower ppm) in the case of CPMG spectra.