Fully automated 3D machine learning model for HPV status characterization in oropharyngeal squamous cell carcinomas based on CT images.

BACKGROUND: Human papillomavirus (HPV) status plays a major role in predicting oropharyngeal squamous cell carcinoma (OPSCC) survival. This study assesses the accuracy of a fully automated 3D convolutional neural network (CNN) in predicting HPV status using CT images. METHODS: Pretreatment CT images from OPSCC patients were used to train a 3D DenseNet-121 model to predict HPV-p16 status. Performance was evaluated by the ROC Curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score. RESULTS: The network achieved a mean AUC of 0.80 ± 0.06. The best-preforming fold had a sensitivity of 0.86 and specificity of 0.92 at the Youden's index. The PPV, NPV, and F1 scores are 0.97, 0.71, and 0.82, respectively. CONCLUSIONS: A fully automated CNN can characterize the HPV status of OPSCC patients with high sensitivity and specificity. Further refinement of this algorithm has the potential to provide a non-invasive tool to guide clinical management.

Comparing accuracy of machine learning approaches to identifying parathyroid adenomas: Lessons and new directions.

PURPOSE: The purpose of this investigation is to understand the accuracy of machine learning techniques to detect biopsy-proven adenomas from similar appearing lymph nodes and factors that influence accuracy by comparing support vector machine (SVM) and bidirectional Long short-term memory (Bi-LSTM) analyses. This will provide greater insight into how these tools could integrate multidimensional data and aid the detection of parathyroid adenomas consistently and accurately. METHODS: Ninety-nine patients were identified; 93 4D-CTs of patients with pathology-proven parathyroid adenomas were reviewed; 94 parathyroid adenomas and 112 lymph nodes were analyzed. A 2D slice through the lesions in each phase was used to perform sequence classification with ResNet50 as the pre-trained network to construct the Bi-LSTM model, and the mean enhancement curves were used to form an SVM model. The model characteristics and accuracy were calculated for the training and validation data sets. RESULTS: On the training data, the area under the curve (AUC) of the Bi-LSTM was 0.99, while the SVM was 0.95 and statistically significant on the DeLong test. The overall accuracy of the Bi-LSTM on the validation data set was 92 %, while the SVM was 88 %. The accuracy for parathyroid adenomas specifically was 93 % for the Bi-LSTM and 83 % for the SVM model. CONCLUSION: Enhancement characteristics are a distinguishing feature that accurately identifies parathyroid adenomas alone. The Bi-LSTM performs statistically better in identifying parathyroid adenomas than the SVM analysis when using both morphologic and enhancement information to distinguish between parathyroid adenomas and lymph nodes. SUMMARY STATEMENT: The Bi-LSTM more accurately identifies parathyroid adenomas than the SVM analysis, which uses both morphologic and enhancement information to distinguish between parathyroid adenomas and lymph nodes, performs statistically better.

Multiparametric machine learning algorithm for human papillomavirus status and survival prediction in oropharyngeal cancer patients.

BACKGROUND: Human papillomavirus (HPV) status influences prognosis in oropharyngeal cancer (OPC). Identifying high-risk patients are critical to improving treatment. We aim to provide a noninvasive opportunity for managing OPC patients by training multiple machine learning pipelines to determine the best model for characterizing HPV status and survival. METHODS: Multi-parametric algorithms were designed using a 492 OPC patient database. HPV status incorporated age, sex, smoking/drinking habits, cancer subsite, TNM, and AJCC 7th edition staging. Survival considered HPV model inputs plus HPV status. Patients were split 4:1 training: testing. Algorithm efficacy was assessed through accuracy and area under the receiver operator characteristic curve (AUC). RESULTS: From 31 HPV status models, ensemble yielded 0.83 AUC and 78.7% accuracy. From 38 survival models, ensemble yielded 0.91 AUC and 87.7% accuracy. CONCLUSION: Results reinforce artificial intelligence's potential to use tumor imaging and patient characterizations for HPV status and outcome prediction. Utilizing these algorithms can optimize clinical guidance and patient care noninvasively.

Quantification of spinal bone marrow fat fraction using three-material decomposition technique on dual-energy CT: A phantom study.

BACKGROUND: Two-material decomposition is insufficient to quantify the fat fraction of spinal bone marrow, which is comprised of a mixture of bone minerals, water, and yellow marrow (fat). PURPOSE: To develop an accurate three-material decomposition-based bone marrow fat fraction ( F F 3 M D $F{F_{3MD}}$ ) quantification technique for dual-energy CT. METHODS: Bone marrow edema phantoms containing trabecular bone minerals, water, and fat were constructed using fat fractions and bone mineral density values matching those expected in healthy and edematous bone, and scanned on a commercial dual-energy CT. Fat quantified by F F 3 M D $F{F_{3MD}}$ were compared to MRI-based fat fraction ( F F M R I $F{F_{MRI}}$ ) and conventional two-material-decomposition-based fat fraction ( F F 2 M D $F{F_{2MD}}$ ) to evaluate its accuracy and dependency on various bone mineral densities. RESULTS: F F 3 M D $F{F_{3MD}}$ demonstrated an excellent correlation with F F M R I $F{F_{MRI}}\;$ (r = 0.97, R2  = 0.96) in the phantom, significantly more accurate than FF2MD when confounding bone minerals are present (50 mg/cm3 : r = 1.02, R2  = 0.95 vs. r = 0.65, R2  = 0.79 (p < 0.01); 100 mg/cm3 : r = 0.81, R2  = 0.47 vs. r = 0.21, R2  = 0.21 (p < 0.05)). CONCLUSIONS: F F 3 M D $F{F_{3MD}}$ accurately quantified bone marrow fat fraction, when compared with F F M R I $F{F_{MRI}}$ , in the specially constructed bone marrow phantom.

Fat Fraction Measurements Using a Three-Material Decomposition Dual-Energy CT Technique Accounting for Bone Minerals: Evaluation in a Bone Marrow Phantom Using MRI as Reference.

Conventional two-material dual-energy CT (DECT) decomposition is insufficient to model bone marrow, which contains three materials: bone minerals, red marrow (water), and yellow marrow (fat). We explore an image-domain three-material decomposition DECT technique accounting for bone minerals in a bone-water-fat phantom. Three-material decomposition fat fraction (FF3MD) exhibited stronger correlation than two-material decomposition fat fraction (FF2MD) with MRI-based fat fraction (r = 0.95 vs r = 0.69). With increasing bone minerals, correlation of FF3MD remained stable (r = 0.81-1.02), whereas correlation of FF2MD decreased (r = 0.21-0.65).

Quantitative assessment of the maxillary sinusitis using computed tomography texture analysis: odontogenic vs non-odontogenic etiology.

OBJECTIVES: The purpose of this study was to investigate computed tomography (CT) texture features of mucosal thickening of maxillary sinus mucosa to differentiate odontogenic maxillary sinusitis (OMS) from non-odontogenic maxillary sinusitis (NOMS). METHODS: Eighteen OMS patients and age- and gender-matched 18 NOMS patients who underwent sinus CT were retrospectively reviewed. OMS patients were identified by histopathological examination of tissues excised at surgery combined with CT imaging findings. Patients with mucosal thickening in the maxillary sinus without apical periodontitis or advanced periodontal bone loss near the maxillary sinus on CT were defined as NOMS. Patients with thin mucosal thickening (< 10 mm), cyst, tumor, post-operative deformity, severe metal artifact precluding visualization of the maxillary sinus, and age younger than 20 years were excluded. CT texture features of the mucosal thickening were analyzed using an in-house developed Matlab-based texture analysis program. Forty-five texture features were extracted from each segmented volume. The results were tested with the Mann-Whitney U test. RESULTS: Six histogram features (mean, median, standard deviation, entropy, geometric mean, harmonic mean) and two gray-level co-occurrence matrix features (entropy, correlation) showed significant differences between OMS and NOMS patients. CONCLUSIONS: CT texture analysis revealed the quantitative differences between OMS and NOMS. The texture features can serve as a quantitative indicator of maxillary sinusitis to differentiate between OMS and NOMS and help prevent incorrect treatment choices.

Prediction of the local treatment outcome in patients with oropharyngeal squamous cell carcinoma using deep learning analysis of pretreatment FDG-PET images.

BACKGROUND: This study aimed to assess the utility of deep learning analysis using pretreatment FDG-PET images to predict local treatment outcome in oropharyngeal squamous cell carcinoma (OPSCC) patients. METHODS: One hundred fifty-four OPSCC patients who received pretreatment FDG-PET were included and divided into training (n = 102) and test (n = 52) sets. The diagnosis of local failure and local progression-free survival (PFS) rates were obtained from patient medical records. In deep learning analyses, axial and coronal images were assessed by three different architectures (AlexNet, GoogLeNET, and ResNet). In the training set, FDG-PET images were analyzed after the data augmentation process for the diagnostic model creation. A multivariate clinical model was also created using a binomial logistic regression model from a patient's clinical characteristics. The test data set was subsequently analyzed for confirmation of diagnostic accuracy. Assessment of local PFS rates was also performed. RESULTS: Training sessions were successfully performed with an accuracy of 74-89%. ROC curve analyses revealed an AUC of 0.61-0.85 by the deep learning model in the test set, whereas it was 0.62 by T-stage, 0.59 by clinical stage, and 0.74 by a multivariate clinical model. The highest AUC (0.85) was obtained with deep learning analysis of ResNet architecture. Cox proportional hazards regression analysis revealed deep learning-based classification by a multivariate clinical model (P < .05), and ResNet (P < .001) was a significant predictor of the treatment outcome. In the Kaplan-Meier analysis, the deep learning-based classification divided the patient's local PFS rate better than the T-stage, clinical stage, and a multivariate clinical model. CONCLUSIONS: Deep learning-based diagnostic model with FDG-PET images indicated its possibility to predict local treatment outcomes in OPSCCs.

Cystic cervical lymph nodes of papillary thyroid carcinoma, tuberculosis and human papillomavirus positive oropharyngeal squamous cell carcinoma: utility of deep learning in their differentiation on CT.

OBJECTIVES: Cervical lymph nodes with internal cystic changes are seen with several pathologies, including papillary thyroid carcinoma (PTC), tuberculosis (TB), and HPV-positive oropharyngeal squamous cell carcinoma (HPV+OPSCC). Differentiating these lymph nodes is difficult in the absence of a known primary tumor or reliable medical history. In this study, we assessed the utility of deep learning in differentiating the pathologic lymph nodes of PTC, TB, and HPV+OPSCC on CT. METHODS: A total of 173 lymph nodes (55 PTC, 58 TB, and 60 HPV+OPSCC) were selected based on pathology records and suspicious morphological features. These lymph nodes were divided into the training set (n = 131) and the test set (n = 42). In deep learning analysis, JPEG lymph node images were extracted from the CT slice that included the largest area of each node and fed into a deep learning training session to create a diagnostic model. Transfer learning was used with the deep learning model architecture of ResNet-101. Using the test set, the diagnostic performance of the deep learning model was compared against the histopathological diagnosis and to the diagnostic performances of two board-certified neuroradiologists. RESULTS: Diagnostic accuracy of the deep learning model was 0.76 (=32/42), whereas those of Radiologist 1 and Radiologist 2 were 0.48 (=20/42) and 0.41 (=17/42), respectively. Deep learning derived diagnostic accuracy was significantly higher than both of the two neuroradiologists (P < 0.01, respectively). CONCLUSION: Deep learning algorithm holds promise to become a useful diagnostic support tool in interpreting cervical lymphadenopathy.

Cystic cervical lymph nodes of papillary thyroid carcinoma, tuberculosis and human papillomavirus positive oropharyngeal squamous cell carcinoma: Comparative CT analysis for their differentiation.

PURPOSE: Cervical lymph nodes with cystic changes are an important finding seen with several pathologies including papillary thyroid carcinoma (PTC), tuberculosis (TB) and HPV-positive oropharyngeal squamous cell carcinoma (HPV + OPSCC). In the absence of known primary tumor or conclusive medical history, differentiating among these nodes is difficult. We compared the pathologic nodes of PTC, TB and HPV + OPSCC to identify imaging features useful for their differentiation. MATERIALS AND METHODS: Fifty-five PTC, 58 TB and 51 HPV + OPSCC nodes were selected based on surgical pathology records and suspicious morphological features. These nodes were compared for morphological features: long axis length, nodal shape, nodal location, presence of cystic change, area of cystic change:area of entire node ratio, Hounsfield unit of the cystic component, degree of enhancement, enhancement pattern, presence of calcification, presence of perinodal infiltration, and presence of surrounding inflammatory changes. RESULTS: PTC nodes formed calcifications more frequently and demonstrated greater enhancement (P < 0.01). TB nodes were characterized by their irregular shape (P < 0.05), irregular enhancement surrounding the cystic change (P < 0.01), greater frequencies of perinodal infiltration (P < 0.01) and surrounding inflammatory changes (P < 0.01). While no unique features were seen with HPV+OPSCC, they were characterized by the absence of those features that distinguished the other groups: these nodes tended to have smooth, circumscribed margins with no hyperenhancement, calcifications or inflammatory changes. PTC and TB nodes were more frequently identified in the lower neck, while HPV+OPSCC nodes were localized to the upper neck (P < 0.01). CONCLUSIONS: PTC, TB and HPV + OPSCC lymph nodes can be differentiated based on their morphologies and locations.

Risk assessment of osteoradionecrosis associated with periodontitis using 18F-FDG PET/CT.

PURPOSE: Osteoradionecrosis (ORN) is a serious complication after radiotherapy (RT), even in the era of intensity modulated radiation therapy (IMRT). The purpose of this study was to evaluate whether 18F-FDG PET/CT can predict ORN associated with periodontal disease in patients with oropharyngeal or oral cavity squamous cell carcinoma (OP/OC SCC) undergoing RT. METHODS: One hundred and five OP/OC SCC patients treated with RT who underwent pretreatment 18F-FDG PET/CT between October 2007 and June 2016 were retrospectively reviewed. A post-treatment diagnosis of ORN was made clinically based on presence of exposed irradiated mandibular bone that failed to heal after a period of three months without persistent or recurrent tumor. The maximum standardized uptake value (SUVmax) of periodontal regions identified on PET/CT was measured for all patients. Image-based staging of periodontitis was also performed using American Academy of Periodontology staging system on CT. RESULTS: Among 105 patients, 14 (13.3 %) developed ORN. The SUVmax of the periodontal region in patients with ORN (3.35 ±â€¯1.23) was significantly higher than patients without ORN (1.92 ±â€¯0.66) (P <  .01). The corresponding CT stage of periodontitis in patients with ORN was significantly higher (2.71±0.47) than patients without ORN (1.80±0.73) (P <  .01). ROC analysis revealed the cut-off values of developing ORN were 2.1 in SUVmax, and II in CT stage of periodontitis. The corresponding AUC was 0.86 and 0.82, respectively. CONCLUSIONS: Pretreatment 18F-FDG PET/CT identification of periodontitis may be helpful to predict the future development of ORN in patients with OP/OC SCC undergoing RT.

Deep learning analysis using FDG-PET to predict treatment outcome in patients with oral cavity squamous cell carcinoma.

OBJECTIVE: To assess the utility of deep learning analysis using 18F-fluorodeoxyglucose (FDG) uptake by positron emission tomography (PET/CT) to predict disease-free survival (DFS) in patients with oral cavity squamous cell carcinoma (OCSCC). METHODS: One hundred thirteen patients with OCSCC who received pretreatment FDG-PET/CT were included. They were divided into training (83 patients) and test (30 patients) sets. The diagnosis of treatment control/failure and the DFS rate were obtained from patients' medical records. In deep learning analyses, three planes of axial, coronal, and sagittal FDG-PET images were assessed by ResNet-101 architecture. In the training set, image analysis was performed for the diagnostic model creation. The test data set was subsequently analyzed for confirmation of diagnostic accuracy. T-stage, clinical stage, and conventional FDG-PET parameters (the maximum and mean standardized uptake value (SUVmax and SUVmean), heterogeneity index, metabolic tumor volume (MTV), and total lesion glycolysis (TLG) were also assessed with determining the optimal cutoff from training dataset and then validated their diagnostic ability from test dataset. RESULTS: In dividing into patients with treatment control and failure, the highest diagnostic accuracy of 0.8 was obtained using deep learning classification, with a sensitivity of 0.8, specificity of 0.8, positive predictive value of 0.89, and negative predictive value of 0.67. In the Kaplan-Meier analysis, the DFS rate was significantly different only with the analysis of deep learning-based classification (p < .01). CONCLUSIONS: Deep learning-based diagnosis with FDG-PET images may predict treatment outcome in patients with OCSCC. KEY POINTS: • Deep learning-based diagnosis of FDG-PET images showed the highest diagnostic accuracy to predict the treatment outcome in patients with oral cavity squamous cell carcinoma. • Deep learning-based diagnosis was shown to differentiate patients between good and poor disease-free survival more clearly than conventional T-stage, clinical stage, and conventional FDG-PET-based parameters.

Amyloidosis in the head and neck: CT findings with clinicopathological correlation.

PURPOSE: To characterize the CT imaging features of head and neck amyloidosis and correlate with extent of disease and clinical outcomes. MATERIALS AND METHODS: This retrospective study included 80 patients with head and neck amyloidosis who underwent soft tissue neck CT imaging between November 2003 and April 2018. The CT imaging features including lesion distribution, morphology (focal, diffuse/circumferential, or combined), presence and pattern of calcification, (punctate or diffuse), and thickness of airway lesion were evaluated and compared with the extent of amyloidosis (localized or systemic), and clinical course (stable, no recurrence, or progression requiring repeated surgical treatment). RESULTS: Localized disease (83.8%, 67/80) was most common with AL type (97.6%, 41/42) representing nearly all cases of head and neck amyloidosis. The larynx was the most frequently affected organ (60.0%, 48/80), specifically the glottis (43.8%, 35/80). Calcification was seen in 65.0% of cases (52/80). Non-airway or tongue lesions were significantly associated with systemic (92.3%, 12/13) as opposed to localized amyloidosis (4.5%, 3/67; P < 0.001). Repeated surgical treatment was significantly associated with laryngeal amyloidosis (35.3%, 12/34; P = 0.002) and multi-centric disease (33.3%, 10/30; P = 0.048). Airway wall thickness in patients who required repeated surgical treatment was significantly greater than in patients with stable or no recurrent disease (P = 0.016). CONCLUSION: Knowledge of the imaging features of head and neck amyloidosis can aid the diagnosis, disease monitoring, and prediction of patients requiring repeated surgical intervention.

Prediction of the human papillomavirus status in patients with oropharyngeal squamous cell carcinoma by FDG-PET imaging dataset using deep learning analysis: A hypothesis-generating study.

PURPOSE: To assess the diagnostic accuracy of imaging-based deep learning analysis to differentiate between human papillomavirus (HPV) positive and negative oropharyngeal squamous cell carcinomas (OPSCCs) using FDG-PET images. METHODS: One hundred and twenty patients with OPSCC who underwent pretreatment FDG-PET/CT were included and divided into the training 90 patients and validation 30 patients cohorts. In the training session, 2160 FDG-PET images were analyzed after data augmentation process by a deep learning technique to create a diagnostic model to discriminate between HPV-positive and HPV-negative OPSCCs. Validation cohort data were subsequently analyzed for confirmation of diagnostic accuracy in determining HPV status by the deep learning-based diagnosis model. In addition, two radiologists evaluated the validation cohort image-data to determine the HPV status based on each tumor's imaging findings. RESULTS: In deep learning analysis with training session, the diagnostic model using training dataset was successfully created. In the validation session, the deep learning diagnostic model revealed sensitivity of 0.83, specificity of 0.83, positive predictive value of 0.88, negative predictive value of 0.77, and diagnostic accuracy of 0.83, while the visual assessment by two radiologists revealed 0.78, 0.5, 0.7, 0.6, and 0.67 (reader 1), and 0.56, 0.67, 0.71, 0.5, and 0.6 (reader 2), respectively. Chi square test showed a significant difference between deep learning- and radiologist-based diagnostic accuracy (reader 1: P = 0.016, reader 2: P = 0.008). CONCLUSIONS: Deep learning diagnostic model with FDG-PET imaging data can be useful as one of supportive tools to determine the HPV status in patients with OPSCC.

Using CT texture analysis to differentiate cystic and cystic-appearing odontogenic lesions.

PURPOSE: Cystic and cystic-appearing odontogenic lesions of the jaw may appear similar on CT imaging. Accurate diagnosis is often difficult although the relationship of the lesion to the tooth root or crown may offer a clue to the etiology. The purpose of this study was to evaluate CT texture analysis as an aid in differentiating cystic and cystic-appearing odontogenic lesions of the jaw. METHODS: This was an IRB-approved retrospective study including 42 pathology-proven dentigerous cysts, 37 odontogenic keratocysts, and 19 ameloblastomas. Each lesion was manually segmented on axial CT images, and textural features were analyzed using an in-house-developed Matlab-based texture analysis program that extracted 47 texture features from each segmented volume. Statistical analysis was performed comparing all pairs of the three types of lesions. RESULTS: Pairwise analysis revealed that nine histogram features, one GLCM feature, three GLRL features, two Laws features, four GLGM features and two Chi-square features showed significant differences between dentigerous cysts and odontogenic keratocysts. Four histogram features and one Chi-square feature showed significant differences between odontogenic keratocysts and ameloblastomas. Two histogram features showed significant differences between dentigerous cysts and ameloblastomas. CONCLUSIONS: CT texture analysis may be useful as a noninvasive method to obtain additional quantitative information to differentiate cystic and cystic-appearing odontogenic lesions of the jaw.

Craniofacial Manifestations of Systemic Disorders: CT and MR Imaging Findings and Imaging Approach.

Many systemic diseases or conditions can affect the maxillofacial bones; however, they are often overlooked or incidentally found at routine brain or head and neck imaging performed for other reasons. Early identification of some conditions may significantly affect patient care and alter outcomes. Early recognition of nonneoplastic hematologic disorders, such as thalassemia and sickle cell disease, may help initiate earlier treatment and prevent serious complications. The management of neoplastic diseases such as lymphoma, leukemia, or Langerhans cell histiocytosis may be different if diagnosed early, and metastases to the maxillofacial bones may be the first manifestation of an otherwise occult neoplasm. Endocrinologic and metabolic disorders also may manifest with maxillofacial conditions. Earlier recognition of osteoporosis may alter treatment and prevent complications such as insufficiency fractures, and identification of acromegaly may lead to surgical treatment if there is an underlying growth hormone-producing adenoma. Bone dysplasias sometimes are associated with skull base foraminal narrowing and subsequent involvement of the cranial nerves. Inflammatory processes such as rheumatoid arthritis and sarcoidosis may affect the maxillofacial bones, skull base, and temporomandibular joints. Radiologists should be familiar with the maxillofacial computed tomographic and magnetic resonance imaging findings of common systemic disorders because these may be the first manifestations of an otherwise unrevealed systemic process with potential for serious complications. Online supplemental material is available for this article. ©RSNA, 2018.