This study evaluated whether deep learning frameworks trained in large datasets can help non-dermatologist physicians improve their accuracy in categorizing the seven most common pigmented skin lesions. Open-source skin images were downloaded from the International Skin Imaging Collaboration (ISIC) archive. Different deep neural networks (DNNs) (n = 8) were trained based on a random dataset constituted of 8015 images. A test set of 2003 images was used to assess the classifiers’ performance at low (300 × 224 RGB) and high (600 × 450 RGB) image resolution and aggregated data (age, sex and lesion localization). We also organized two different contests to compare the DNN performance to that of general practitioners by means of unassisted image observation. Both at low and high image resolution, the DNN framework differentiated dermatological images with appreciable performance. In all cases, the accuracy was improved when adding clinical data to the framework. Finally, the least accurate DNN outperformed general practitioners. The physician’s accuracy was statistically improved when allowed to use the output of this algorithmic framework as guidance. DNNs are proven to be high performers as skin lesion classifiers and can improve general practitioner diagnosis accuracy in a routine clinical scenario.

Artificial intelligence can be a key tool in the context of assisting in the diagnosis of dermatological conditions, particularly when performed by general practitioners with limited or no access to high resolution optical equipment. This study evaluates the performance of deep convolutional neural networks (DNNs) in the classification of seven pigmented skin lesions. Additionally, it assesses the improvement ratio in the classification performance when utilized by general practitioners. Open-source skin images were downloaded from the ISIC archive. Different DNNs (n=8) were trained based on a random dataset constituted by 8,015 images. A test set of 2,003 images has been used to assess the classifiers performance at low (300 x 224 RGB) and high (600 x 450 RGB) image resolution and aggregated clinical data (age, sex and lesion localization). We have also organized two different contests to compare the DNNs performance to that of general practitioners by means of unassisted image observation. Both at low and high image resolution, the DNNs framework being trained differentiated dermatological images with appreciable performance. In all cases, accuracy has been improved when adding clinical data to the framework. Finally, the lowest accurate DNN outperformed general practitioners. Physicians accuracy was statistically improved when allowed to use the output of this algorithmic framework as guidance. DNNS are proven to be high performers as skin lesion classifiers. The aim is to include these AI tools in the context of general practitioners whilst improving their diagnosis accuracy in a routine clinical scenario when or where the use of high-resolution equipment is not accessible.

e15526 Background: Practice guidelines recommend using P to treat K-Ras WT mCRC patients where it was shown to significantly extend overall survival (OS). Still, a proportion of patients will not achieve this goal. We propose a simplified predictive score to identify patients who are likely to benefit from P treatment. Methods: NCT00364013 was used as training dataset (TRS) (n = 460) with NCT00339183 (TES1) (n = 479) and NCT00113763 (TES2) (n = 191) as validations sets. Datasets were obtained from www.projectdatasphere.org and included K-Ras WT mCRC patients treated with P in combination or not with FOLFOX4 (FOL) or FOLFIRI as 1st, 2nd, or 3rd line therapy. TRS was used to generate synthetic representations (SRs) for each patient through the integration of 36 clinical and analytical features collected, respectively, during the screening phase and the first month of inclusion. These SRs were then input into a deep learning framework (DLF) to identify subgroup of patients based on their similarities. The resultant subpopulations were correlated with OS. Differential variables between subgroups were identified through feature contribution analysis and included in a multivariable logistic regression model. Independent predictive factors found to be statistically significant were used to generate a predictive score of P response at baseline that was validated in the test sets. Results: DLF identified two different subpopulations: SPA (n = 162) and SPB (n = 298). Patients in SPA had a lower risk of death when treated with P/FOL compared to FOL (HR 0.68 95%CI 0.48-0.99; p = 0.04). Patients in SPB showed no significant differences between P/FOL and FOL (p = 0.27). Feature contribution analysis identified 15 differential features between both subpopulations. From these, CEA > 174 ng/ml, ALP > 131 IU, LDH > 703 IU, and platelets > 374 109/L were selected to create a simplified predictive score for P response ranging 0-18 (if > than the depicted values: 6.5 points for CEA, 5.5 for LDH, and 3 points for each other characteristic). When applied to TRS, this score yielded an area under the curve of 0.87 (95%CI: 0.84-0.91). A score ≥8.5 was positively correlated to a longer OS after P/FOL compared to FOL (HR 0.65 95%CI 0.43-0.98; p = .04). No significant differences were observed between P/FOL and FOL in patients with a score < 8.5 (p = 0.89). The predictive score was then validated in the two test sets with similar results (score ≥8.5, TES1: HR 0.59 95%CI 0.40-0.88 p = .009; TES2: HR: HR 0.58 95%CI 0.35-0.96 p = .03; score < 8.5, TES1: p = .5; TES2: p = .1). Conclusions: Based on CEA, ALP, LDH and platelet baseline levels, this easily applicable predictive score might be helpful to accurately select K-Ras WT mCRC patients who would benefit from P treatment. Further work is required to validate this approach in prospective cohorts of patients.