Retrace Labs

Ali Sadat

Vasant Kearney

Hamid Hekmatian

Stephen Chan

Wenxiang Deng

Ashwini Jha

Ming Ted Wong

Hamid Hakmatian

Yash Patel

A novel GAN is trained to predict high fidelity synthetic images based on low quality input dental images. The GAN further takes input anatomic masks as inputs with each image, the masks labeling pixels of the image corresponding to dental features. The GAN includes an encoder-decoder generator with semantically aware normalization between stages of the decoder according to the masks. The predicted synthetic dental image and an unpaired dental image are evaluated by a first discriminator of the GAN to obtain a realism estimate. The synthetic image and an unpaired dental image may be processed using a pretrained dental encoder to obtain a perceptual loss. The GAN is trained with the realism estimate, perceptual loss, and L1 loss. Utilization may include inputting noisy, low contrast, low resolution, blurry, or degraded dental images and outputting high resolution, denoised, high contrast, deobfuscated, and sharp dental images.

Generative adversarial network for dental image super-resolution, image sharpening, and denoising

In medicine and dentistry, image quality affects computer vision accuracy. However, some problems are more tolerant of noise depending on disease severity and radiographic obviousness. There is a need to have a noise estimation model that adapts to each specific domain. A noise estimation model is trained to output a set of domain noise estimates for an input image, each estimate indicating an impact of noise present in the input image on a particular domain, e.g. labeling of a dental feature such as a dental anatomy, pathology, or treatment. The noise estimation model is trained by processing image pairs with a set of machine learning models for a plurality of domains, the image pairs including a raw image and a modified image obtained by adding noise to the raw image. Outputs of the set of machine learning models for the raw and modified images are compared to obtain measured noise metrics. The noise estimation model processes the modified image and is trained to estimate noise metrics. The noise estimation model is modified according to differences between the measured noise metrics and estimated noise metrics.

Dental image quality prediction platform using domain specific artificial intelligence

A GAN is trained to process input images and produce a synthetic dental image. The GAN further takes masks as inputs with each image, the masks labeling pixels of the image corresponding to dental features (anatomy and/or treatments). The GAN includes an encoder-decoder with normalization between stages of the decoder according to the masks. A synthetic image and an unpaired dental image is evaluated by a first discriminator of the GAN to obtain a realism estimate. The synthetic image and an unpaired dental image may be processed using a pretrained dental encoder to obtain a perceptual loss. The GAN is trained with the realism estimate and perceptual loss. Utilization may include modifying a mask for an input image to include or exclude a shape of a feature such that the synthetic image includes or excludes a dental feature.

Dental image synthesis using generative adversarial networks with semantic activation blocks

A comprehensive dental readiness platform is presented. Dental patient data including an image, proposed treatments, and a dental form are received and processed by first machine learning models to obtain clinical findings and predicted values for fields of the dental form. The clinical findings and other results are processed by a second machine learning model to obtain predictions of a future dental condition of a patient. The second machine learning model utilizes an ensemble of Transformer Neural Networks, Long-Short-Term-Memory Networks, Convolutional Neural Networks, and Tree-Based Algorithms to predict the dental readiness classification, dental readiness durability, dental readiness error, dental emergency likelihood, prognosis, and alternative treatment options.

Artificial intelligence platform for determining dental readiness

Dental images are processed according to a first machine learning model to determine teeth labels. The teeth labels and image are concatenated and processed using a second machine learning model to label anatomy including CEJ, JE, GM, and Bone. The anatomy labels, teeth labels, and image are concatenated and processed using a third machine learning model to obtain feature measurements, such as pocket depth and clinical attachment level. The feature measurements, anatomy labels, teeth labels, and image may be concatenated and input to a fourth machine learning model to obtain a diagnosis for a periodontal condition. Feature measurements and/or the diagnosis may be processed according to a diagnosis hierarchy to determine whether a treatment is appropriate. Machine learning models may further be used to reorient, decontaminate, and restore the image prior to processing. Machine learning models may be embodied as CNN, GAN, and cyclic GAN.

Artificial intelligence architecture for identification of periodontal features

Training a generator includes processing a dental image using the generator to obtain a synthetic pathology label, such has a pixel mask indicating portions of the dental image representing caries. The synthetic pathology label is compared to a target pathology label for the dental image and the generator is updated according to the comparison. The synthetic pathology may be evaluated by a discriminator along with a real pathology label to obtain a realism estimate. The discriminator and generator may be updated according to accuracy of the realism estimate. Inputs to the generator may further include tooth labels and/or labels of restorations. Machine learning models may be trained to label restorations and defects in restorations. A machine learning model may be trained to identify the surface of a tooth having a pathology thereon.

System and methods for restorative dentistry treatment planning using adversarial learning

Dental images are processed according to a first machine learning model to determine teeth labels. The teeth labels and image are processed using a second machine learning model to label anatomy. The anatomy labels, teeth labels, and image are processed using a third machine learning model to obtain feature measurements, such as pocket depth and clinical attachment level. The feature measurements, labels, and image may be input to a fourth machine learning model to obtain a diagnosis for a periodontal condition. Machine learning models may further be used to reorient, decontaminate, and restore the image prior to processing. A machine learning model may be trained with images and randomly generated masks in order to perform inpainting of dental images with missing information.

Inpainting dental images with missing anatomy

A first machine learning model is trained to classify dental anatomy and/or pathologies represented in an input dental image or to generate a label (pixel mask) for dental anatomy and/or pathologies represented in the input dental image. A final layer, such as one of two fully connected layers, may be removed from the first machine learning model to obtain a modified machine learning model. Hidden features output from the modified machine learning model may be input to a LSTM model that outputs a text sequence. The LSTM model may be trained with images labeled with text sequences to output a text sequence for a given input dental image.

Systems and method for artificial-intelligence-based dental image to text generation

A machine learning model is trained to predict pixel spacing, distance, and volumetric measurements. Training images are obtained by inpainting around an original image and scaling the inpainted image to obtain the training image having a different pixel spacing than the original image. The machine learning model may include an encoder, a transformer, a first TC layer, and a second TC layer. During training, loss may be obtained from a comparison of the output to the first TC layer to a coarse pixel spacing matrix and a comparison of the output of the second TC layer to a fine pixel spacing matrix. During utilization, the pixel spacing of an image may be obtained using the machine learning model and used to correct the image or measurements obtained from the image.

Growing community of inventors

San Francisco, CA, United States of America

Ali Sadat