ABSTRACT
In this paper, we propose a novel method for the detection of small lesions in digital medical images. Our approach is based on a multi-context ensemble of convolutional neural networks (CNNs), aiming at learning different levels of image spatial context and improving detection performance. The main innovation behind the proposed method is the use of multiple-depth CNNs, individually trained on image patches of different dimensions and then combined together. In this way, the final ensemble is able to find and locate abnormalities on the images by exploiting both the local features and the surrounding context of a lesion. Experiments were focused on two well-known medical detection problems that have been recently faced with CNNs: microcalcification detection on full-field digital mammograms and microaneurysm detection on ocular fundus images. To this end, we used two publicly available datasets, INbreast and E-ophtha. Statistically significantly better detection performance were obtained by the proposed ensemble with respect to other approaches in the literature, demonstrating its effectiveness in the detection of small abnormalities.
Subject(s)
Breast Neoplasms/diagnostic imaging , Fundus Oculi , Image Processing, Computer-Assisted/methods , Neural Networks, Computer , Breast Neoplasms/diagnosis , Breast Neoplasms/pathology , Deep Learning , Humans , Mammography/methodsABSTRACT
PURPOSE: The aim of this study was to assess how often women with undetected calcifications in prior screening mammograms are subsequently diagnosed with invasive cancer. METHODS: From a screening cohort of 63,895 women, exams were collected from 59,690 women without any abnormalities, 744 women with a screen-detected cancer and a prior negative exam, 781 women with a false positive exam based on calcifications, and 413 women with an interval cancer. A radiologist identified cancer-related calcifications, selected by a computer-aided detection system, on mammograms taken prior to screen-detected or interval cancer diagnoses. Using this ground truth and the pathology reports, the sensitivity for calcification detection and the proportion of lesions with visible calcifications that developed into invasive cancer were determined. RESULTS: The screening sensitivity for calcifications was 45.5%, at a specificity of 99.5%. A total of 68.4% (n = 177) of cancer-related calcifications that could have been detected earlier were associated with invasive cancer when diagnosed. CONCLUSIONS: Screening sensitivity for detection of malignant calcifications is low. Improving the detection of these early signs of cancer is important, because the majority of lesions with detectable calcifications that are not recalled immediately but detected as interval cancer or in the next screening round are invasive at the time of diagnosis.
Subject(s)
Breast Neoplasms/complications , Calcinosis/diagnosis , Early Diagnosis , Adult , Aged , Breast Neoplasms/pathology , Calcinosis/complications , Calcinosis/pathology , Female , Humans , Mammography , Mass Screening , Middle AgedABSTRACT
The present study aims to evaluate and validate the Italian version of Khalfa's questionnaire on hyperacusis (HQ). We recruited 117 patients (64 men, 53 women, mean age 53 years, range 14-88) with tinnitus for at least 3 months as a primary disorder. All patients completed the THI and the Italian version of the HQ and underwent audiometry, pitch and loudness tinnitus matching, otoacoustic emissions with distortion products (DPOAE) and uncomfortable loudness level (ULL). The overall performance of the tests was evaluated and compared using the area under the RO C curve (AUC) relative to the tests. The cut-off of the HQ was calculated. We also assessed the Cronbach's alpha (αC) for the HQ and its three major dimensions (attentional - αC1, emotional - αC2 and social - αC3). Statistical analysis showed no correlation between DPOAE, audiometry, ULL and gender. We observed a high correlation (p < 0.05) between hyperacusis and ULL described by the Spearman's ρ index (rs = 0.72). We found a cut-off of 16 indicative of hyperacusis comparing the area under the RO C curve (AUC) of HQ and audiometry, taken as a diagnostic reference, (sensitivity = 67.9% and specificity = 72.2%). The reliability of HQ was confirmed by a high αC = 0.89. The αC for the single dimensional scales were, respectively, αC1 = 0.73, αC2 = 0.72 and αC3 = 0.81. The Italian version of the HQ is recommended for proper and complete classification of patients with tinnitus and hyperacusis. From our study, we found a cut-off of 16 instead of the cut-off of 28 described as very high by other authors. Moreover, ULL was an important variable and can be discriminating in the evaluation of hyperacusis.
Subject(s)
Hyperacusis/diagnosis , Self Report , Tinnitus/diagnosis , Adolescent , Adult , Aged , Aged, 80 and over , Female , Humans , Italy , Language , Male , Middle Aged , Young AdultABSTRACT
Finding abnormalities in diagnostic images is a difficult task even for expert radiologists because the normal tissue locations largely outnumber those with suspicious signs which may thus be missed or incorrectly interpreted. For the same reason the design of a Computer-Aided Detection (CADe) system is very complex because the large predominance of normal samples in the training data may hamper the ability of the classifier to recognize the abnormalities on the images. In this paper we present a novel approach for computer-aided detection which faces the class imbalance with a cascade of boosting classifiers where each node is trained by a learning algorithm based on ranking instead of classification error. Such approach is used to design a system (CasCADe) for the automated detection of clustered microcalcifications (µCs), which is a severely unbalanced classification problem because of the vast majority of image locations where no µC is present. The proposed approach was evaluated with a dataset of 1599 full-field digital mammograms from 560 cases and compared favorably with the Hologic R2CAD ImageChecker, one of the most widespread commercial CADe systems. In particular, at the same lesion sensitivity of R2CAD (90%) on biopsy proven malignant cases, CasCADe and R2CAD detected 0.13 and 0.21 false positives per image (FPpi), respectively (p-value=0.09), whereas at the same FPpi of R2CAD (0.21), CasCADe and R2CAD detected 93% and 90% of true lesions respectively (p-value=0.11) thus showing that CasCADe can compete with high-end CADe commercial systems.
Subject(s)
Breast Neoplasms/diagnostic imaging , Calcinosis/diagnostic imaging , Mammography/methods , Radiographic Image Enhancement/methods , Radiographic Image Interpretation, Computer-Assisted/methods , Algorithms , Female , HumansABSTRACT
BACKGROUND AND AIMS: Subsite-specific incidence rates of colorectal cancer (CRC) and adenomas may vary considerably by race, sex and age as well as due to different screening strategies. We assessed variations in the anatomical distribution of adenomas according to age and sex in an average-risk screening cohort testing positive at immunological faecal occult blood test (i-FOBT) in northern Italy. METHODS: Data from 2,281 consecutive asymptomatic i-FOBT-positive subjects ageing 50-70 years undergone colonoscopy were reviewed. Size, number, macroscopic and histological features of all adenomas found as well as their proximal or distal location in relation to the splenic flexure were examined. Odds ratios (OR) of proximal neoplasms, according to the presence of distal neoplasms and other selected covariates were assessed by multiple logistic regression analysis. RESULTS: A total of 2,599 neoplasms were found in 1,396 patients. Of these, 116 (5 %) were colorectal cancers, diagnosed in 106 patients. Out of 2,483 adenomas found, 1,564 (63 %) were sessile, 795 (32 %) were peduncolated and 124 (5 %) were flat-type; 54 % of all adenomas were tubular, 36 % were tubulovillous or villous, and 10 % were serrated adenomas. The majority of neoplasms (66 %) were located in the distal colon. Tumour subsite distribution was consistent in both sexes, whereas significant proximal migration of neoplasms occurred in the older age cohort. Indeed, the rate of proximal neoplasms in patients aged ≥60 years was 37 % as compared with 29 % in those ageing 50-59 years. Male gender (OR 1.84), age of 60 years or older (OR 1.44), having a family history of colorectal neoplasms (OR 1.47) and presence of at least 1 distal advanced adenoma (OR 1.63) were all significant predictors of advanced proximal neoplasms. CONCLUSIONS: A left to right shift of colorectal adenomas with increasing age is evident in northern Italian asymptomatic i-FOBT-positive population. Advanced proximal neoplasms are not uncommon in subjects with or without distal adenomas, especially after 60 years of age. This should be carefully considered when implementing public screening strategies for CRC since the use of flexible sigmoidoscopy as a screening tool, particularly in older age groups, appears to be less effective.
Subject(s)
Adenoma/pathology , Colorectal Neoplasms/pathology , Early Detection of Cancer/methods , Hematologic Tests/methods , Occult Blood , Sex Characteristics , Adenoma/blood , Adenoma/diagnosis , Adenoma/epidemiology , Age Factors , Aged , Cohort Studies , Colonoscopy , Colorectal Neoplasms/blood , Colorectal Neoplasms/diagnosis , Colorectal Neoplasms/epidemiology , Female , Humans , Italy/epidemiology , Male , Middle Aged , Risk FactorsABSTRACT
BACKGROUND: The results of a pilot colorectal cancer screening programme by biennial immunochemical faecal occult blood test (FOBT) are reported. METHODS: All residents aged between 50 and 69 years in the Italian province of Lecco were invited to have a FOBT. Those with a positive result were offered colonoscopy. FOBT uptake and compliance with colonoscopy were assessed. Detection rate and positive predictive value (PPV) for cancer and adenoma were calculated. Tumour stages were compared between screen-detected cancers and other colorectal cancers diagnosed within the target age group. RESULTS: Some 38,693 (49.6 per cent) of 78,083 individuals had a FOBT and 2392 (6.2 per cent) had a positive result. Colorectal cancer was diagnosed in 4.6 per cent and advanced adenoma in 32.7 per cent. PPVs were 4.0 per cent for cancer, 28.1 per cent for advanced adenoma and 36.6 per cent for any adenoma. There was a significant difference in incidence of stage III/IV disease between screened and non-screened cohorts. Compliance for colonoscopy was 92.0 per cent. Major determinants of compliance were age less than 59 years, female sex, high education level and non-manual work. CONCLUSION: These results justify extension of colorectal cancer screening to other regions of Italy.
Subject(s)
Colorectal Neoplasms/diagnosis , Early Detection of Cancer , Mass Screening/methods , Occult Blood , Aged , Aged, 80 and over , Colonoscopy/statistics & numerical data , Educational Status , Female , Humans , Italy , Male , Middle Aged , Patient Compliance/statistics & numerical data , Pilot Projects , Sex DistributionABSTRACT
Criteria for evaluating the classification reliability of a neural classifier and for accordingly making a reject option are proposed. Such an option, implemented by means of two rules which can be applied independently of topology, size, and training algorithms of the neural classifier, allows one to improve the classification reliability. It is assumed that a performance function P is defined which, taking into account the requirements of the particular application, evaluates the quality of the classification in terms of recognition, misclassification, and reject rates. Under this assumption the optimal reject threshold value, determining the best trade-off between reject rate and misclassification rate, is the one for which the function P reaches its absolute maximum. No constraints are imposed on the form of P, but the ones necessary in order that P actually measures the quality of the classification process. The reject threshold is evaluated on the basis of some statistical distributions characterizing the behavior of the classifier when operating without reject option; these distributions are computed once the training phase of the net has been completed. The method has been tested with a neural classifier devised for handprinted and multifont printed characters, by using a database of about 300000 samples. Experimental results are discussed.