Radiology: AI may improve rads detection of hepatic cancer
Artificially intelligent (AI) neural networks demonstrate strong sensitivity and specificity in the detection of liver cancer using PET/CT, while also improving the accuracy of interpreting radiologists who took the networks’ findings into account, suggesting an important future adjunct role for the systems in cancer detection, according to a study published in the March issue of Radiology.
Artificial neural networks (also called parallel distributed processors) are mathematical algorithms that synthesize input data to produce solutions or predictions. In the present study, two neural networks, a lesion-independent and a lesion-dependent network, were fed information on patient demographics and physiologic and anatomic PET/CT scan data among patients imaged for intrahepatic cancer (including standardized uptake values (SUVs), blood glucose levels and other data). The lesion-dependent network also incorporated lesion SUV and lesion SUV standard deviation.
Two radiologists and both networks independently interpreted 98 PET/CT whole-body scans for intrahepatic lesion malignancy or benignity. Patients were injected with 18 fluorodeoxyglucose (FDG) prior to PET/CT and were imaged using MR, which served as the reference standard for all interpretations. The networks and both reviewers provided estimates of the probability of hepatic malignancy, between 0.00 and 1.00, indicated to the hundredths place.
The area under the curve (AUC) and the receiver operating characteristic curve for the two blinded radiologists were 0.786 and 0.796. Both the lesion-independent and lesion-dependent networks performed favorably compared with the radiologists, demonstrating AUCs of 0.896 and 0.905, respectively.
The researchers re-read all PET/CT images, incorporating the networks’ estimates. Reviewer one’s AUC improved from 0.786 to 0.924 and reviewer two’s from 0.796 to 0.881; both improvements reached significance. These figures demonstrated highly significant correlation to MR results.
“With each network, 10 patients were identified correctly as having metastatic disease despite the absence of visually apparent FDG abnormalities, suggesting that, at least in certain cases, the network may have interpretive capacity that a human observer lacks,” wrote Ori Preis, MD, and co-authors from the department of radiology at Massachusetts General Hospital in Boston.
The authors attributed the smaller improvement in reader two’s AUC after incorporating network data to reduced trust in the computer system. Whereas observer one determined his unblinded value by averaging his blinded estimate with that of the networks, the authors said observer two adjusted his estimates more cautiously.
From these varying approaches, the authors inferred that the “methodology of individual readers in clinical practice may vary from the methodology of these two readers during incorporation of results with the network, and it would therefore be difficult to predict the alteration in interpretive accuracy for future readers.”
Preis and co-authors said that although the lesion-dependent network performed superiorly to the independent network, the difference was “very small, suggesting that the actual lesion data are not central to the methodology of the neural network in interpretation.”
The authors acknowledged the flaws of MRI in detecting intrahepatic lesions, while affirming the modality as the best available to radiologists at present and highly accurate in the present study. In one case, however, radiologists identified what they believed was a metastasis on MRI, which biopsy proved to be a hemangioma. The neural networks did not identify this lesion as metastatic.
“It is clear from these preliminary results that networks can allow the correct interpretation of certain PET scans, which a human observer may misinterpret, and vice versa, owing to the different interpretative algorithms applied by the neural networks,” the authors said.
They continued, “It is our opinion that a neural network may serve as a clinical adjunct to aid in interpretation.”
Artificial neural networks (also called parallel distributed processors) are mathematical algorithms that synthesize input data to produce solutions or predictions. In the present study, two neural networks, a lesion-independent and a lesion-dependent network, were fed information on patient demographics and physiologic and anatomic PET/CT scan data among patients imaged for intrahepatic cancer (including standardized uptake values (SUVs), blood glucose levels and other data). The lesion-dependent network also incorporated lesion SUV and lesion SUV standard deviation.
Two radiologists and both networks independently interpreted 98 PET/CT whole-body scans for intrahepatic lesion malignancy or benignity. Patients were injected with 18 fluorodeoxyglucose (FDG) prior to PET/CT and were imaged using MR, which served as the reference standard for all interpretations. The networks and both reviewers provided estimates of the probability of hepatic malignancy, between 0.00 and 1.00, indicated to the hundredths place.
The area under the curve (AUC) and the receiver operating characteristic curve for the two blinded radiologists were 0.786 and 0.796. Both the lesion-independent and lesion-dependent networks performed favorably compared with the radiologists, demonstrating AUCs of 0.896 and 0.905, respectively.
The researchers re-read all PET/CT images, incorporating the networks’ estimates. Reviewer one’s AUC improved from 0.786 to 0.924 and reviewer two’s from 0.796 to 0.881; both improvements reached significance. These figures demonstrated highly significant correlation to MR results.
“With each network, 10 patients were identified correctly as having metastatic disease despite the absence of visually apparent FDG abnormalities, suggesting that, at least in certain cases, the network may have interpretive capacity that a human observer lacks,” wrote Ori Preis, MD, and co-authors from the department of radiology at Massachusetts General Hospital in Boston.
The authors attributed the smaller improvement in reader two’s AUC after incorporating network data to reduced trust in the computer system. Whereas observer one determined his unblinded value by averaging his blinded estimate with that of the networks, the authors said observer two adjusted his estimates more cautiously.
From these varying approaches, the authors inferred that the “methodology of individual readers in clinical practice may vary from the methodology of these two readers during incorporation of results with the network, and it would therefore be difficult to predict the alteration in interpretive accuracy for future readers.”
Preis and co-authors said that although the lesion-dependent network performed superiorly to the independent network, the difference was “very small, suggesting that the actual lesion data are not central to the methodology of the neural network in interpretation.”
The authors acknowledged the flaws of MRI in detecting intrahepatic lesions, while affirming the modality as the best available to radiologists at present and highly accurate in the present study. In one case, however, radiologists identified what they believed was a metastasis on MRI, which biopsy proved to be a hemangioma. The neural networks did not identify this lesion as metastatic.
“It is clear from these preliminary results that networks can allow the correct interpretation of certain PET scans, which a human observer may misinterpret, and vice versa, owing to the different interpretative algorithms applied by the neural networks,” the authors said.
They continued, “It is our opinion that a neural network may serve as a clinical adjunct to aid in interpretation.”